Acknowledgements: The authors thank Paul Invernizzi for helpful discussions.
This research was supported by the chair Risques Financiers , Fondation du Risque, and by the BNP Paribas chair Futures of Quantitative Finance .
Introduction
Motivation
In financial modeling, as stressed by [33 ] , verifying the martingale property of the asset price is often a first sanity check: when the discounted price is a strict local martingale, several well-known pathologies can appear, including violations of put–call parity and distortions in the asymptotic behavior of implied volatility wings [16 , 26 , 27 ] . To prove martingality, beyond Novikov’s criterion, which only applies to rather simple cases, the standard approach is to use a change-of-measure argument that reduces the problem to the study of explosion for an auxiliary SDE, as in [30 Chapter 3.7] . Such an approach is used by [34 ] , [29 ] , and [28 ] , which study, at different levels of generality, the martingale property and moment integrability of “lognormal volatility models” of the form
∀ t ∈ [ 0 , + ∞ ) , d S t S t = e X t d W t ρ , \forall t \in[0,+\infty), \qquad\frac{dS_t}{S_t}=e^{X_t}dW_t^\rho, ∀ t ∈ [ 0 , + ∞ ) , S t d S t = e X t d W t ρ ,
X t = x 0 e − λ t + μ λ ( 1 − e − λ t ) + α ∫ 0 t e − λ ( t − s ) d W s . (1) X_t=x_0e^{-\lambda t}+\frac{\mu}{\lambda}\left(1-e^{-\lambda t}\right)+\alpha\int_0^t e^{-\lambda(t-s)}dW_s. \tag*{(1)} X t = x 0 e − λ t + λ μ ( 1 − e − λ t ) + α ∫ 0 t e − λ ( t − s ) d W s . (1)
Here we take x 0 , μ ∈ R x_0,\mu\in\mathbb{R} x 0 , μ ∈ R and λ , α > 0 \lambda,\alpha>0 λ , α > 0 . We define the Brownian motion W ρ W^\rho W ρ by
∀ t ≥ 0 , W t ρ : = ρ W t + 1 − ρ 2 W t ⊥ , with ρ ∈ [ − 1 , 1 ] . \forall t\geq0,\qquad W_t^\rho:=\rho W_t+\sqrt{1-\rho^2}W_t^\perp,\qquad\text{with }\rho\in[-1,1]. ∀ t ≥ 0 , W t ρ := ρ W t + 1 − ρ 2 W t ⊥ , with ρ ∈ [ − 1 , 1 ] .
Here W W W and W ⊥ W^\perp W ⊥ are two independent Brownian motions. Then W ρ W^\rho W ρ is a Brownian motion whose instantaneous correlation with W W W is equal to ρ \rho ρ . The authors show that the martingale property holds only when ρ ≤ 0 \rho\leq0 ρ ≤ 0 . Later, [8 ] obtained more general martingality criteria for SDE-based models. In non-Markovian frameworks, [20 ] studies a class of fractional stochastic volatility models, including the rough Bergomi model, of the form
∀ t ∈ [ 0 , + ∞ ) , d S t S t = σ ( t , X t ) d W t ρ , X t = ∫ 0 t K ( t − s ) d W s . (2) \forall t\in[0,+\infty),\qquad\frac{dS_t}{S_t}=\sigma(t,X_t)dW_t^\rho,\qquad X_t=\int_0^t K(t-s)dW_s. \tag*{(2)} ∀ t ∈ [ 0 , + ∞ ) , S t d S t = σ ( t , X t ) d W t ρ , X t = ∫ 0 t K ( t − s ) d W s . (2)
Under the coefficient and kernel assumptions in [20 ] , ρ < 0 \rho<0 ρ < 0 gives martingality. For ρ > 0 \rho>0 ρ > 0 , a sufficient growth condition on σ \sigma σ , as in [20 Theorem 1(2), Eq. (1.3)] , gives strict local martingality when K K K dominates the rough kernel defined by
∀ τ ∈ ( 0 , + ∞ ) , ∀ H ∈ ( 0 , 1 2 ) , K H ( τ ) : = τ H − 1 2 Γ ( H + 1 2 ) . (3) \forall\tau\in(0,+\infty),\ \forall H\in\left(0,\frac{1}{2}\right),\qquad K_H(\tau):=\frac{\tau^{H-\frac{1}{2}}}{\Gamma\left(H+\frac{1}{2}\right)}. \tag*{(3)} ∀ τ ∈ ( 0 , + ∞ ) , ∀ H ∈ ( 0 , 2 1 ) , K H ( τ ) := Γ ( H + 2 1 ) τ H − 2 1 . (3)
Here Γ \Gamma Γ denotes the Euler gamma function. Adding a linear drift to the volatility factor in (2) gives the extension
∀ t ∈ [ 0 , + ∞ ) , d S t S t = σ ( t , X t ) d W t ρ , X t = β ( t ) + κ ∫ 0 t K ( t − s ) X s d s + ν ∫ 0 t K ( t − s ) d W s . (4) \forall t\in[0,+\infty),\qquad\frac{dS_t}{S_t}=\sigma(t,X_t)dW_t^\rho,\qquad X_t=\beta(t)+\kappa\int_0^t K(t-s)X_s\,ds+\nu\int_0^t K(t-s)dW_s. \tag*{(4)} ∀ t ∈ [ 0 , + ∞ ) , S t d S t = σ ( t , X t ) d W t ρ , X t = β ( t ) + κ ∫ 0 t K ( t − s ) X s d s + ν ∫ 0 t K ( t − s ) d W s . (4)
Here β : R + → R \beta:\mathbb{R}_+\to\mathbb{R} β : R + → R is continuous and deterministic, κ ∈ R \kappa\in\mathbb{R} κ ∈ R , ν ∈ R ∖ { 0 } \nu\in\mathbb{R}\setminus\{0\} ν ∈ R ∖ { 0 } , and K K K is locally square-integrable. This framework includes the Volterra–Stein–Stein model studied in [4 ] , when σ ( t , x ) = x \sigma(t,x)=x σ ( t , x ) = x or σ ( t , x ) = ∣ x ∣ \sigma(t,x)=|x| σ ( t , x ) = ∣ x ∣ , while the case κ = 0 \kappa=0 κ = 0 recovers Volterra volatility models of Bergomi type. In particular, with an exponential specification of σ \sigma σ , one recovers rough/Volterra Bergomi-type models, including the original Bergomi model and, more generally, exponential Ornstein–Uhlenbeck (exp-OU) models, whereas a polynomial specification gives, for instance, the Quintic model introduced in [3 ] . For the exponential Bergomi specifications below, we take κ = 0 \kappa=0 κ = 0 , ν > 0 \nu>0 ν > 0 , and σ ( t , x ) = φ ( t ) e x \sigma(t,x)=\varphi(t)e^x σ ( t , x ) = φ ( t ) e x , where φ : R + → ( 0 , ∞ ) \varphi:\mathbb{R}_+\to(0,\infty) φ : R + → ( 0 , ∞ ) is continuous. For the Quintic specification, one uses σ ( t , x ) = φ ( t ) ( a 0 + a 1 x + a 3 x 3 + a 5 x 5 ) \sigma(t,x)=\varphi(t)(a_0+a_1x+a_3x^3+a_5x^5) σ ( t , x ) = φ ( t ) ( a 0 + a 1 x + a 3 x 3 + a 5 x 5 ) , with a 0 , a 1 , a 3 ≥ 0 a_0,a_1,a_3\geq0 a 0 , a 1 , a 3 ≥ 0 and a 5 > 0 a_5>0 a 5 > 0 , and an exponential kernel in (4) with κ = 0 \kappa=0 κ = 0 and ν > 0 \nu>0 ν > 0 .
A large literature on rough volatility has developed in recent years to capture both the strong path-dependence of volatility and its strong irregularity in time. The paper [21 ] shows empirically that log-volatility behaves, across a wide range of time scales, like a fractional process with Hurst parameter H H H close to 0.1, far below the Brownian value H = 1 2 H = \frac{1}{2} H = 2 1 . This provides an empirical motivation for modeling volatility through rough Volterra-type dynamics rather than through classical Markovian diffusions.
In the rough Heston model of El Euch and Rosenbaum [18 ] , the variance satisfies a Volterra equation with a fractional kernel. The model retains the square-root diffusion coefficient of the classical Heston model:
∀ t ∈ [ 0 , + ∞ ) , d S t S t = X t d W t ρ , \forall t \in[0,+\infty), \qquad\frac{\mathrm{d}S_t}{S_t} = \sqrt{X_t}\,\mathrm{d}W_t^\rho, ∀ t ∈ [ 0 , + ∞ ) , S t d S t = X t d W t ρ ,
X t = X 0 + λ ∫ 0 t K H ( t − s ) ( θ − X s ) d s + ν ∫ 0 t K H ( t − s ) X s d W s , (5) X_t = X_0 + \lambda\int_0^t K_H(t-s)(\theta-X_s)\,\mathrm{d}s + \nu\int_0^t K_H(t-s)\sqrt{X_s}\,\mathrm{d}W_s, \tag*{(5)} X t = X 0 + λ ∫ 0 t K H ( t − s ) ( θ − X s ) d s + ν ∫ 0 t K H ( t − s ) X s d W s , (5)
where H ∈ ( 0 , 1 2 ) H \in(0,\frac{1}{2}) H ∈ ( 0 , 2 1 ) and K H K_H K H is defined by (3) , X 0 ≥ 0 X_0 \ge0 X 0 ≥ 0 , and λ , θ , ν > 0 \lambda,\theta,\nu> 0 λ , θ , ν > 0 . The martingality of the rough Heston asset price was proved in [2 ] for all correlations ρ \rho ρ .
A related line of research concerns quadratic rough models. Gatheral, Jusselin and Rosenbaum [22 ] introduced such a model to address the joint calibration of SPX and VIX smiles, showing that rough volatility combined with an explicit price-feedback (Zumbach) effect can overcome limitations of more classical continuous-path models. Dandapani, Jusselin and Rosenbaum [17 ] provided a microstructural foundation for this class of models through quadratic Hawkes processes, while Rosenbaum and Zhang [32 ] developed an efficient calibration method based on multi-factor approximations and deep neural networks. Using a resolvent, as proposed in [13 ] , we incorporate a linear mean-reversion term into the kernel. In the fractional case, the transformed kernel is a Mittag–Leffler kernel. This gives the following parametrization. We use the convention in which the sign of the Brownian driver in the asset price equation is absorbed into the correlation parameter ρ \rho ρ . We define the function V V V by
∀ x ∈ R , V ( x ) : = a 0 + a 1 ( x − a 2 ) 2 , a 0 ≥ 0 , a 1 > 0 , a 2 ∈ R . \forall x \in\mathbb{R}, \qquad V(x) := a_0 + a_1(x-a_2)^2,\qquad a_0 \ge0,\quad a_1 > 0,\quad a_2 \in\mathbb{R}. ∀ x ∈ R , V ( x ) := a 0 + a 1 ( x − a 2 ) 2 , a 0 ≥ 0 , a 1 > 0 , a 2 ∈ R .
We set
∀ t ∈ [ 0 , + ∞ ) , d S t S t = V ( X t ) d W t ρ , X t = β ( t ) + ∫ − ∞ t K ( t − s ) V ( X s ) d W s . (6) \forall t \in[0,+\infty), \qquad\frac{\mathrm{d}S_t}{S_t} = \sqrt{V(X_t)}\,\mathrm{d}W_t^\rho,\qquad X_t = \beta(t) + \int_{-\infty}^t K(t-s)\sqrt{V(X_s)}\,\mathrm{d}W_s. \tag*{(6)} ∀ t ∈ [ 0 , + ∞ ) , S t d S t = V ( X t ) d W t ρ , X t = β ( t ) + ∫ − ∞ t K ( t − s ) V ( X s ) d W s . (6)
where K K K denotes the possibly transformed Volterra kernel and β \beta β the corresponding input curve, chosen to be constant in [13 ] . Integrals over negative times use a two-sided Brownian motion and a prescribed history of X X X . By a two-sided Brownian motion W = ( W t ) t ∈ R W=(W_t)_{t\in\mathbb{R}} W = ( W t ) t ∈ R , we mean a continuous process with W 0 = 0 W_0=0 W 0 = 0 such that ( W t ) t ≥ 0 (W_t)_{t\geq0} ( W t ) t ≥ 0 and ( W − t ) t ≥ 0 (W_{-t})_{t\geq0} ( W − t ) t ≥ 0 are independent standard Brownian motions. These history integrals are assumed to be well defined and to yield continuous initial curves. A minimal extension, the Quadratic Volterra Heston+ model of [14 ] , adds a nonnegative term to V V V for x < x 0 x<x_0 x < x 0 by defining
∀ x ∈ R , V ^ ( x ) : = V ( x ) + b 0 ( ( x 0 − x ) + ) p . (7) \forall x \in\mathbb{R}, \qquad\widehat{V}(x) := V(x) + b_0\bigl((x_0-x)^+\bigr)^p. \tag*{(7)} ∀ x ∈ R , V ( x ) := V ( x ) + b 0 ( ( x 0 − x ) + ) p . (7)
with b 0 ≥ 0 b_0 \geq0 b 0 ≥ 0 , x 0 ∈ R x_0 \in\mathbb{R} x 0 ∈ R , and p > 0 p > 0 p > 0 , where x + : = x ∨ 0 x^+ := x \vee0 x + := x ∨ 0 and x − : = ( − x ) ∨ 0 x^- := (-x) \vee0 x − := ( − x ) ∨ 0 . The model is given by
∀ t ∈ [ 0 , + ∞ ) , d S t S t = V ^ ( X t ) d W t ρ , X t = β ( t ) + ∫ − ∞ t K ( t − s ) V ^ ( X s ) d W s . (8) \forall t \in[0,+\infty), \qquad\frac{\mathrm{d}S_t}{S_t} = \sqrt{\widehat{V}(X_t)}\,\mathrm{d}W_t^\rho,\qquad X_t = \beta(t) + \int_{-\infty}^t K(t-s)\sqrt{\widehat{V}(X_s)}\,\mathrm{d}W_s. \tag*{(8)} ∀ t ∈ [ 0 , + ∞ ) , S t d S t = V ( X t ) d W t ρ , X t = β ( t ) + ∫ − ∞ t K ( t − s ) V ( X s ) d W s . (8)
When b 0 = 0 b_0=0 b 0 = 0 , (8) reduces to (6) . For b 0 = 0 b_0=0 b 0 = 0 , a 0 > 0 a_0>0 a 0 > 0 , and ρ = 1 \rho=1 ρ = 1 , strict local martingality in the Markovian case is already noted in [9 p. 2] . The extension proved here concerns the smooth Volterra kernels in Assumption 4.2 below.
Recent studies in quantitative finance, such as [38 ] , [39 ] , [15 ] , [11 ] , [25 ] and [7 ] , have highlighted the significant path-dependence of volatility, which is largely explained by past asset returns—rather than by the current asset price alone. These studies show that the instantaneous volatility of financial markets is mostly endogenous and depends only marginally on additional random factors. Motivated by an empirical investigation showing that a large fraction of the variations of an equity volatility index, such as the VIX, can be explained by past daily returns and past squared daily returns , Guyon and Lekeufack [25 ] introduced the class of path-dependent volatility (PDV) models, where the instantaneous volatility depends explicitly on two factors: a (signed) return factor R 1 , t R_{1,t} R 1 , t and a variance factor R 2 , t R_{2,t} R 2 , t , built from past returns and past squared returns weighted by memory kernels, respectively. A convenient formulation is the autonomous SVE system of [6 ] . We consider an asset price ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 evolving for t ≥ 0 t\geq0 t ≥ 0 with instantaneous volatility σ t = X t \sigma_t=X_t σ t = X t according to
∀ t ∈ [ 0 , + ∞ ) , d S t S t = X t d W t , X t = β 0 + β 1 R 1 , t + β 2 R 2 , t , R 1 , t = ∫ − Δ t K 1 ( t − s ) X s d W s , R 2 , t = ∫ − Δ t K 2 ( t − s ) X s 2 d s , (9) \begin{aligned}
&\forall t\in[0,+\infty),\qquad\frac{\mathrm{d}S_t}{S_t}=X_t\,\mathrm{d}W_t,\\
&X_t=\beta_0+\beta_1R_{1,t}+\beta_2\sqrt{R_{2,t}},\\
&R_{1,t}=\int_{-\Delta}^{t}K_1(t-s)X_s\,\mathrm{d}W_s,\\
&R_{2,t}=\int_{-\Delta}^{t}K_2(t-s)X_s^2\,\mathrm{d}s, \tag*{(9)}
\end{aligned} ∀ t ∈ [ 0 , + ∞ ) , S t d S t = X t d W t , X t = β 0 + β 1 R 1 , t + β 2 R 2 , t , R 1 , t = ∫ − Δ t K 1 ( t − s ) X s d W s , R 2 , t = ∫ − Δ t K 2 ( t − s ) X s 2 d s , (9)
where β 0 ≥ 0 \beta_0\geq0 β 0 ≥ 0 , Δ ∈ [ 0 , + ∞ ] \Delta\in[0,+\infty] Δ ∈ [ 0 , + ∞ ] , and the nonnegative convolution kernels K 1 , K 2 : ( 0 , ∞ ) → R + K_1,K_2:(0,\infty)\to\mathbb{R}_+ K 1 , K 2 : ( 0 , ∞ ) → R + weight past returns and past squared returns, respectively. Throughout the paper, the PDV parameters satisfy β 1 < 0 \beta_1<0 β 1 < 0 and β 2 > 0 \beta_2>0 β 2 > 0 .
The negative-time path of the two-sided Brownian motion W W W , independent of its positive-time increments, encodes the history of the model before time zero, namely the returns observed before time 0. When K 1 K_1 K 1 and K 2 K_2 K 2 each have two exponential components, the model admits a continuous-time four-factor PDV (4FPDV) representation: R 1 R_1 R 1 and R 2 R_2 R 2 can each be represented by two factors with different decay rates. Nutz and Riveros Valdevenito [31 ] prove existence and pathwise uniqueness of global strong solutions for the 4FPDV model. In a more general (non-Markovian) Volterra-kernel setting, [6 ] gives existence and uniqueness results for the Guyon–Lekeufack PDV dynamics with general kernels. Gazzani and Guyon [23 ] give a method for pricing and calibration in the 4FPDV model. They add an “upside” term to generate sufficiently high volatility when R 1 , t > 0 R_{1,t}>0 R 1 , t > 0 and improve the fit of implied volatility smiles. More precisely, they consider
∀ t ∈ [ 0 , + ∞ ) , X t = β 0 + F 1 ( R 1 , t ) + β 2 R 2 , t , ∀ y ∈ R , F 1 ( y ) : = β 1 y + β 1 , 2 y 2 1 { y > 0 } . (10) \begin{aligned}
&\forall t\in[0,+\infty),\qquad X_t=\beta_0+F_1(R_{1,t})+\beta_2\sqrt{R_{2,t}},\\
&\forall y\in\mathbb{R},\qquad F_1(y):=\beta_1y+\beta_{1,2}y^2\mathbf{1}_{\{y>0\}}. \tag*{(10)}
\end{aligned} ∀ t ∈ [ 0 , + ∞ ) , X t = β 0 + F 1 ( R 1 , t ) + β 2 R 2 , t , ∀ y ∈ R , F 1 ( y ) := β 1 y + β 1 , 2 y 2 1 { y > 0 } . (10)
Here we assume β 1 , 2 > 0 \beta_{1,2}>0 β 1 , 2 > 0 . We call the models obtained by allowing in (10) a function F 1 F_1 F 1 that is nondecreasing for sufficiently large arguments upside-enhanced Volterra–Guyon–Lekeufack models . A related extension is the polynomial PDV model, studied in a working paper. We define its radicand and instantaneous volatility by
∀ t ∈ [ 0 , + ∞ ) , X t : = β 0 + β 1 R 1 , t + β 1 ( 2 ) R 1 , t 2 + β 2 R 2 , t , σ t : = ( X t ) + . (11) \forall t\in[0,+\infty),\qquad X_t:=\beta_0+\beta_1R_{1,t}+\beta_1^{(2)}R_{1,t}^2+\beta_2R_{2,t},\qquad\sigma_t:=\sqrt{(X_t)_+}. \tag*{(11)} ∀ t ∈ [ 0 , + ∞ ) , X t := β 0 + β 1 R 1 , t + β 1 ( 2 ) R 1 , t 2 + β 2 R 2 , t , σ t := ( X t ) + . (11)
Here we have β 1 ( 2 ) > 0 \beta_1^{(2)}>0 β 1 ( 2 ) > 0 , and we assume that
∀ T ∈ ( 0 , + ∞ ) , P 0 ( ∀ t ∈ [ 0 , T ] , X t ≥ 0 ) = 1. (12) \forall T\in(0,+\infty),\qquad\mathbb{P}_0\left(\forall t\in[0,T],\ X_t\geq0\right)=1. \tag*{(12)} ∀ T ∈ ( 0 , + ∞ ) , P 0 ( ∀ t ∈ [ 0 , T ] , X t ≥ 0 ) = 1. (12)
A sufficient condition for (12) is β 0 + β 1 y + β 1 ( 2 ) y 2 ≥ 0 \beta_0+\beta_1y+\beta_1^{(2)}y^2\geq0 β 0 + β 1 y + β 1 ( 2 ) y 2 ≥ 0 for every y ∈ R y\in\mathbb{R} y ∈ R , since β 2 > 0 \beta_2>0 β 2 > 0 and R 2 , t ≥ 0 R_{2,t}\geq0 R 2 , t ≥ 0 . For all PDV models considered below, we assume that σ 0 > 0 \sigma_0>0 σ 0 > 0 . Within the family of path-dependent volatility (PDV) models introduced in [25 ] , [31 ] analyzes the martingale property in a two-factor PDV setting and shows that the asset price is a true martingale as long as the volatility remains bounded from below. One contribution of this article is to extend this result to more general kernels and to show that, for the Volterra–Guyon–Lekeufack model (9) , writing σ t \sigma_t σ t for the instantaneous volatility, under the conditional law P 0 \mathbb{P}_0 P 0 defined in § 2.1 , for every T > 0 T > 0 T > 0 ,
P 0 ( ∀ t ∈ [ 0 , T ] , σ t ≥ 0 ) = 1 ⟹ ( S t ) t ∈ [ 0 , T ] is a martingale under P 0 . \mathbb{P}_0\left(\forall t \in[0,T],\ \sigma_t \ge0\right) = 1 \quad\Longrightarrow\quad(S_t)_{t\in[0,T]} \text{ is a martingale under } \mathbb{P}_0. P 0 ( ∀ t ∈ [ 0 , T ] , σ t ≥ 0 ) = 1 ⟹ ( S t ) t ∈ [ 0 , T ] is a martingale under P 0 .
Recently, Abi Jaber, Gassiat and Sotnikov [5 ] proved sharp martingality criteria for finite-order signature volatility models. Their results cover models of polynomial type and extend to signatures driven by Ornstein–Uhlenbeck processes, including the Quintic OU model. For Brownian signatures truncated at order N ≥ 2 N \ge2 N ≥ 2 and a nonzero coefficient of W t N N ! \frac{W_t^N}{N!} N ! W t N in the volatility, with ρ ≠ 0 \rho\ne0 ρ = 0 , the price is a martingale if and only if N N N is odd and this coefficient has the opposite sign to ρ \rho ρ .
Bondi and Pulido [10 ] give a Feller-type boundary test for scalar stochastic Volterra equations with nonsingular kernels and constant initial input. We study nonlinear functions of coupled Volterra factors with initial curves encoding the past. For smooth kernels, we prove an explosion criterion with multiplicative noise by controlling the memory drift in a semimartingale representation.
Martingality follows either by excluding an explosive direction while controlling the opposite direction, or from one-sided growth bounds on the Girsanov drift. Both criteria allow rough kernels. The second adapts Lyapunov methods for moment control to Volterra factors that need not be semimartingales, using a Volterra chain-rule inequality when Itô’s formula cannot be applied. Related estimates were used by Bourdon and Jeannin [12 ] for subcritical moments in the rough Bergomi model.
The applications of our martingality criteria to several popular Volterra models in mathematical finance are summarized in Table 1 . Below, we consider the following two generalized Volterra–Guyon–Lekeufack specifications, with β 0 = β 0 , 1 \beta_0 = \beta_{0,1} β 0 = β 0 , 1 outside F 1 F_1 F 1 , in which the continuous function F 1 F_1 F 1 is defined, for β 0 , 1 , β 0 , 2 , y 0 ∈ R \beta_{0,1}, \beta_{0,2}, y_0 \in\mathbb{R} β 0 , 1 , β 0 , 2 , y 0 ∈ R and β 1 , 1 < 0 < β 1 , 2 \beta_{1,1} < 0 < \beta_{1,2} β 1 , 1 < 0 < β 1 , 2 , by
Table | Sufficient conditions for martingality and strict local martingality in the listed models. Throughout the table, ρ ≠ 0 \rho\ne0 ρ = 0 . The martingality results allow rough kernels under the assumptions of the results cited in the last column, whereas the strict local martingality results proved here require the smooth-kernel regularity in Assumption 4.2 below, but are strongly expected to extend to rough kernels. The same criteria hold with P \mathbb{P} P replaced throughout by the conditional law P 0 \mathbb{P}_0 P 0 defined in § 2.1 , for almost every initial history.
Table 1. Sufficient conditions for martingality and strict local martingality in the listed models.
∀ y ∈ R , F 1 ( y ) : = ( β 1 , 1 y ) ∨ ( β 0 , 2 − β 0 , 1 + β 1 , 2 y ) , (13) \forall y \in\mathbb{R}, \qquad F_1(y) := (\beta_{1,1}y) \vee(\beta_{0,2} - \beta_{0,1} + \beta_{1,2}y), \tag*{(13)} ∀ y ∈ R , F 1 ( y ) := ( β 1 , 1 y ) ∨ ( β 0 , 2 − β 0 , 1 + β 1 , 2 y ) , (13)
∀ y ∈ R , F 1 ( y ) : = ( β 1 , 1 y ) ∨ ( β 0 , 2 − β 0 , 1 + β 1 , 2 ( y − y 0 ) + ) . (14) \forall y \in\mathbb{R}, \qquad F_1(y) := (\beta_{1,1}y) \vee\left(\beta_{0,2} - \beta_{0,1} + \beta_{1,2}\sqrt{(y-y_0)^+}\right). \tag*{(14)} ∀ y ∈ R , F 1 ( y ) := ( β 1 , 1 y ) ∨ ( β 0 , 2 − β 0 , 1 + β 1 , 2 ( y − y 0 ) + ) . (14)
Framework and main results
A unified Volterra framework
On a filtered probability space ( Ω , F , ( F t ) t ≥ 0 , P ) (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\ge0},\mathbb{P}) ( Ω , F , ( F t ) t ≥ 0 , P ) satisfying the usual conditions, we consider a Brownian motion ( W t ) t ≥ 0 (W_t)_{t\ge0} ( W t ) t ≥ 0 . We assume that a regular conditional probability given F 0 \mathcal{F}_0 F 0 exists. Unless stated otherwise, we work under the regular conditional law defined by P 0 : = P ( ⋅ ∣ F 0 ) \mathbb{P}_0:=\mathbb{P}(\cdot\mid\mathcal{F}_0) P 0 := P ( ⋅ ∣ F 0 ) , and write E 0 \mathbb{E}_0 E 0 for its expectation. We fix a realization of the F 0 \mathcal{F}_0 F 0 -history outside an exceptional P \mathbb{P} P -null set, so that P 0 \mathbb{P}_0 P 0 is a fixed probability measure. All assertions under this law are understood for almost every initial history. For an integrable random variable Z Z Z , E 0 [ Z ∣ F t ] \mathbb{E}_0[Z\mid\mathcal{F}_t] E 0 [ Z ∣ F t ] denotes its conditional expectation under this measure, given the full history up to time t t t . We consider two initial curves G 1 , G 2 : R + → R G_1,G_2:\mathbb{R}_+\to\mathbb{R} G 1 , G 2 : R + → R , an initial price S 0 > 0 S_0>0 S 0 > 0 , and a continuous adapted factor process Y : = ( Y 1 , Y 2 ) Y:=(Y_1,Y_2) Y := ( Y 1 , Y 2 ) solving (PDV). The curves G 1 , G 2 G_1,G_2 G 1 , G 2 , viewed as random elements of C ( R + , R ) C(\mathbb{R}_+,\mathbb{R}) C ( R + , R ) , and S 0 S_0 S 0 are F 0 \mathcal{F}_0 F 0 -measurable. For statements under P \mathbb{P} P , we additionally assume E [ S 0 ] < ∞ \mathbb{E}[S_0]<\infty E [ S 0 ] < ∞ . We enlarge the space, if necessary, by a Brownian motion ( W t ⊥ ) t ≥ 0 (W_t^\perp)_{t\ge0} ( W t ⊥ ) t ≥ 0 independent of the factor system ( W , Y ) (W,Y) ( W , Y ) , and take, for t ≥ 0 t \ge0 t ≥ 0 , F t \mathcal{F}_t F t to be the usual augmentation of
F 0 ∨ Gen ( ( W s ) s ∈ [ 0 , t ] , ( Y s ) s ∈ [ 0 , t ] , ( W s ⊥ ) s ∈ [ 0 , t ] ) . \mathcal{F}_0 \vee\operatorname{Gen}\left((W_s)_{s\in[0,t]},(Y_s)_{s\in[0,t]},(W_s^{\perp})_{s\in[0,t]}\right). F 0 ∨ Gen ( ( W s ) s ∈ [ 0 , t ] , ( Y s ) s ∈ [ 0 , t ] , ( W s ⊥ ) s ∈ [ 0 , t ] ) .
Here Gen ( ⋅ ) \operatorname{Gen}(\cdot) Gen ( ⋅ ) denotes the generated sigma-field. Under P 0 \mathbb{P}_0 P 0 , W W W and W ⊥ W^{\perp} W ⊥ are independent Brownian motions. We consider a discounted asset price ( S t ) t ≥ 0 (S_t)_{t\ge0} ( S t ) t ≥ 0 . For ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] , we set W ρ = ρ W + 1 − ρ 2 W ⊥ W^\rho= \rho W + \sqrt{1-\rho^2}W^{\perp} W ρ = ρ W + 1 − ρ 2 W ⊥ . The price satisfies
∀ t ∈ [ 0 , + ∞ ) , d S t S t = σ t d W t ρ , σ t : = σ ( t , X t ) . \forall t \in[0,+\infty), \qquad\frac{dS_t}{S_t}=\sigma_t\,dW_t^\rho, \qquad\sigma_t:=\sigma(t,X_t). ∀ t ∈ [ 0 , + ∞ ) , S t d S t = σ t d W t ρ , σ t := σ ( t , X t ) .
with
∀ t ∈ [ 0 , + ∞ ) , X t = β ( t ) + F 1 ( Y 1 , t ) + F 2 ( Y 2 , t ) , Y 1 , t = G 1 ( t ) + ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X s ) d W s , Y 2 , t = G 2 ( t ) + ∫ 0 t K 2 ( t − s ) Γ 2 ( s , X s ) d s . (15) \begin{aligned}
\forall t \in[0,+\infty), \qquad X_t &= \beta(t)+F_1(Y_{1,t})+F_2(Y_{2,t}),\\
Y_{1,t} &= G_1(t)+\int_0^t K_1(t-s)\Gamma_1(s,X_s)\,dW_s,\\
Y_{2,t} &= G_2(t)+\int_0^t K_2(t-s)\Gamma_2(s,X_s)\,ds. \tag*{(15)}
\end{aligned} ∀ t ∈ [ 0 , + ∞ ) , X t Y 1 , t Y 2 , t = β ( t ) + F 1 ( Y 1 , t ) + F 2 ( Y 2 , t ) , = G 1 ( t ) + ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X s ) d W s , = G 2 ( t ) + ∫ 0 t K 2 ( t − s ) Γ 2 ( s , X s ) d s . (15)
for deterministic functions
F 1 , F 2 : R → R , Γ 1 , Γ 2 : R + × R → R , σ : R + × R → R , β : R + → R . F_1,F_2:\mathbb{R}\to\mathbb{R}, \qquad\Gamma_1,\Gamma_2:\mathbb{R}_+\times\mathbb{R}\to\mathbb{R}, \qquad\sigma:\mathbb{R}_+\times\mathbb{R}\to\mathbb{R}, \qquad\beta:\mathbb{R}_+\to\mathbb{R}. F 1 , F 2 : R → R , Γ 1 , Γ 2 : R + × R → R , σ : R + × R → R , β : R + → R .
Throughout the paper, the kernels are convolution kernels: K 1 , K 2 : ( 0 , ∞ ) → R + K_1,K_2:(0,\infty)\to\mathbb{R}_+ K 1 , K 2 : ( 0 , ∞ ) → R + are measurable, and we use K i ( t − s ) K_i(t-s) K i ( t − s ) for s < t s<t s < t . We call Y 1 Y_1 Y 1 and Y 2 Y_2 Y 2 the first and second Volterra factors. For the purposes of this paper, we extend the coefficient x ↦ x x\mapsto\sqrt{x} x ↦ x , initially defined on R + \mathbb{R}_+ R + , continuously to R \mathbb{R} R by x ↦ x + x\mapsto\sqrt{x^+} x ↦ x + .
Example 1. This framework includes most of the models considered in the literature. The representation of each model is given in Table 2 . For the first five models, we take G 1 = G 2 = 0 G_1=G_2=0 G 1 = G 2 = 0 . In the Quadratic Volterra Heston+ row, G 1 G_1 G 1 also includes the input curve β \beta β from (8) , so the framework coefficient β \beta β is zero. The first two rows use K 1 = K 2 = K K_1=K_2=K K 1 = K 2 = K .
Table 2. Examples of models covered by the general framework (PDV).
Remark 2.1. This representation is not unique: different choices of the functions F i F_i F i , Γ i \Gamma_i Γ i , and σ \sigma σ may correspond to the same model.
We use the following assumptions only where stated. Kernel regularity is summarized in Table 5 , and polynomial-growth exponents in Table 3 .
Table 3. Polynomial-growth exponents in Assumption 1.2 for functions defined in Table 2.
Table 5. Regularity exponents and resolvent properties of standard kernels.
Assumption 1. The convolution kernels K 1 , K 2 K_1,K_2 K 1 , K 2 are nonnegative and satisfy
∀ t ∈ ( 0 , + ∞ ) , ∀ i ∈ { 1 , 2 } , ∫ 0 t K i ( s ) d s > 0. \forall t\in(0,+\infty),\quad\forall i\in\{1,2\},\quad\int_0^t K_i(s)\,\mathrm{d}s>0. ∀ t ∈ ( 0 , + ∞ ) , ∀ i ∈ { 1 , 2 } , ∫ 0 t K i ( s ) d s > 0.
Increment regularity of the kernels
There exist δ 1 , δ 2 > 0 \delta_1,\delta_2>0 δ 1 , δ 2 > 0 such that, for every T > 0 T>0 T > 0 , the following bounds hold with a constant C T > 0 C_T>0 C T > 0 :
∀ h ∈ [ 0 , T ] , ( ∫ 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 , (16) \forall h\in[0,T],\quad\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*{(16)} ∀ h ∈ [ 0 , T ] , ( ∫ 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 , (16)
∫ 0 T − h ∣ K 2 ( u + h ) − K 2 ( u ) ∣ d u + ∫ 0 h K 2 ( u ) d u ≤ C T h δ 2 . (17) \int_0^{T-h}\lvert K_2(u+h)-K_2(u)\rvert\,\mathrm{d}u+\int_0^h K_2(u)\,\mathrm{d}u\le C_T h^{\delta_2}. \tag*{(17)} ∫ 0 T − h ∣ K 2 ( u + h ) − K 2 ( u )∣ d u + ∫ 0 h K 2 ( u ) d u ≤ C T h δ 2 . (17)
Polynomial growth of F i F_i F i and $\Gamma_i
The functions G i G_i G i , F i F_i F i , Γ i \Gamma_i Γ i , σ \sigma σ , and β \beta β are continuous, and there exist exponents γ i ≥ 0 \gamma_i\ge0 γ i ≥ 0 , i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , f 1 , + , f 1 , − , f 2 > 0 f_{1,+},f_{1,-},f_2>0 f 1 , + , f 1 , − , f 2 > 0 , and a constant C F 2 > 0 C_{F_2}>0 C F 2 > 0 , such that, with f 1 : = f 1 , + ∨ f 1 , − f_1:=f_{1,+}\vee f_{1,-} f 1 := f 1 , + ∨ f 1 , − , γ i f i ≤ 1 \gamma_i f_i\le1 γ i f i ≤ 1 , i ∈ { 1 , 2 } i\in\{1,2\} i ∈ { 1 , 2 } , and for every T > 0 T>0 T > 0 , there exists C T > 0 C_T>0 C T > 0 satisfying
∀ i ∈ { 1 , 2 } , ∀ ( t , x ) ∈ [ 0 , T ] × R , ∣ Γ i ( t , x ) ∣ ≤ C T ( 1 + ∣ x ∣ γ i ) , \forall i\in\{1,2\},\quad\forall(t,x)\in[0,T]\times\mathbb{R},\quad\lvert\Gamma_i(t,x)\rvert\le C_T(1+\lvert x\rvert^{\gamma_i}), ∀ i ∈ { 1 , 2 } , ∀ ( t , x ) ∈ [ 0 , T ] × R , ∣ Γ i ( t , x )∣ ≤ C T ( 1 + ∣ x ∣ γ i ) ,
∀ y ∈ R , ∣ F 2 ( y ) ∣ ≤ C F 2 ( 1 + ∣ y ∣ f 2 ) . \forall y\in\mathbb{R},\quad\lvert F_2(y)\rvert\le C_{F_2}(1+\lvert y\rvert^{f_2}). ∀ y ∈ R , ∣ F 2 ( y )∣ ≤ C F 2 ( 1 + ∣ y ∣ f 2 ) .
and, for each s ∈ { − 1 , 1 } s \in\{-1,1\} s ∈ { − 1 , 1 } , there exists ε s ∈ { − 1 , 1 } \varepsilon_s \in\{-1,1\} ε s ∈ { − 1 , 1 } such that, for some y ∗ > 0 y_* > 0 y ∗ > 0 , F 1 ( s y ) F_1(sy) F 1 ( sy ) has sign ε s \varepsilon_s ε s for y ≥ y ∗ y \ge y_* y ≥ y ∗ and
∃ 0 < c ≤ C < + ∞ such that ∀ y ∈ [ y ∗ , + ∞ ) , c y f 1 , s ≤ ε s F 1 ( s y ) ≤ C y f 1 , s . \exists0 < c \le C < +\infty\text{ such that } \forall y \in[y_*,+\infty), \qquad cy^{f_{1,s}} \le\varepsilon_s F_1(sy) \le Cy^{f_{1,s}}. ∃0 < c ≤ C < + ∞ such that ∀ y ∈ [ y ∗ , + ∞ ) , c y f 1 , s ≤ ε s F 1 ( sy ) ≤ C y f 1 , s .
The quadratic upside specification does not satisfy Assumption 1.2, since γ 1 f 1 = 2 > 1 \gamma_1 f_1 = 2 > 1 γ 1 f 1 = 2 > 1 . The two specifications (A) and (B) do satisfy its growth bounds.
For a bounded stopping time θ \theta θ , we consider
∀ t ∈ [ 0 , + ∞ ) , X ~ t θ = β ( t ) + F 1 ( Y ~ 1 , t θ ) + F 2 ( Y ~ 2 , t θ ) , \forall t \in[0,+\infty), \qquad\widetilde{X}^{\theta}_t = \beta(t) + F_1(\widetilde{Y}^{\theta}_{1,t}) + F_2(\widetilde{Y}^{\theta}_{2,t}), ∀ t ∈ [ 0 , + ∞ ) , X t θ = β ( t ) + F 1 ( Y 1 , t θ ) + F 2 ( Y 2 , t θ ) ,
Y ~ 1 , t θ = G 1 ( t ) + ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X ~ s θ ) d W s + ρ ∫ 0 t 1 { s ≥ θ } K 1 ( t − s ) Γ 1 ( s , X ~ s θ ) σ ( s , X ~ s θ ) d s , \widetilde{Y}^{\theta}_{1,t} = G_1(t) + \int_0^t K_1(t-s)\Gamma_1(s,\widetilde{X}^{\theta}_s)\,\mathrm{d}W_s
+ \rho\int_0^t \mathbf{1}_{\{s\ge\theta\}}K_1(t-s)\Gamma_1(s,\widetilde{X}^{\theta}_s)\sigma(s,\widetilde{X}^{\theta}_s)\,\mathrm{d}s, Y 1 , t θ = G 1 ( t ) + ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X s θ ) d W s + ρ ∫ 0 t 1 { s ≥ θ } K 1 ( t − s ) Γ 1 ( s , X s θ ) σ ( s , X s θ ) d s ,
Y ~ 2 , t θ = G 2 ( t ) + ∫ 0 t K 2 ( t − s ) Γ 2 ( s , X ~ s θ ) d s . \widetilde{Y}^{\theta}_{2,t} = G_2(t) + \int_0^t K_2(t-s)\Gamma_2(s,\widetilde{X}^{\theta}_s)\,\mathrm{d}s. Y 2 , t θ = G 2 ( t ) + ∫ 0 t K 2 ( t − s ) Γ 2 ( s , X s θ ) d s .
We set Y ~ θ : = ( Y ~ 1 θ , Y ~ 2 θ ) \widetilde{Y}^{\theta} := (\widetilde{Y}^{\theta}_1,\widetilde{Y}^{\theta}_2) Y θ := ( Y 1 θ , Y 2 θ ) . In particular, for all t ∈ [ 0 , θ ] t \in[0,\theta] t ∈ [ 0 , θ ] , we have ( X ~ t θ , Y ~ t θ ) = ( X t , Y t ) (\widetilde{X}^{\theta}_t,\widetilde{Y}^{\theta}_t) = (X_t,Y_t) ( X t θ , Y t θ ) = ( X t , Y t ) . By Lemma 3.1 below, martingality of S S S after θ \theta θ is equivalent to non-explosion of X ~ θ \widetilde{X}^{\theta} X θ . For T , n > 0 T,n>0 T , n > 0 , we define the probability measure Q n \mathbb{Q}_n Q n by
d Q n d P 0 ∣ F θ + T : = S θ + ( T ∧ τ ( θ , n ) X ) S θ , \left.\frac{\mathrm{d}\mathbb{Q}_n}{\mathrm{d}\mathbb{P}_0}\right|_{\mathcal{F}_{\theta+T}}
:= \frac{S_{\theta+(T\wedge\tau^{X}_{(\theta,n)})}}{S_\theta}, d P 0 d Q n F θ + T := S θ S θ + ( T ∧ τ ( θ , n ) X ) ,
where τ ( θ , n ) X : = inf { u ≥ 0 : ∣ X θ + u ∣ ≥ n } \tau^{X}_{(\theta,n)} := \inf\{u \ge0 : |X_{\theta+u}| \ge n\} τ ( θ , n ) X := inf { u ≥ 0 : ∣ X θ + u ∣ ≥ n } . (For all hitting times, we use the convention inf ∅ = + ∞ \inf\varnothing= +\infty inf ∅ = + ∞ .) Up to θ + ( T ∧ τ ( θ , n ) X ) \theta+(T\wedge\tau^{X}_{(\theta,n)}) θ + ( T ∧ τ ( θ , n ) X ) , this change of measure adds the following drift to the first Volterra factor:
ρ ∫ 0 t 1 { s ≥ θ } K 1 ( t − s ) Γ 1 ( s , X ~ s θ ) σ ( s , X ~ s θ ) d s . (18) \rho\int_0^t \mathbf{1}_{\{s\ge\theta\}}K_1(t-s)\Gamma_1(s,\widetilde{X}^{\theta}_s)\sigma(s,\widetilde{X}^{\theta}_s)\,\mathrm{d}s. \tag*{(18)} ρ ∫ 0 t 1 { s ≥ θ } K 1 ( t − s ) Γ 1 ( s , X s θ ) σ ( s , X s θ ) d s . (18)
The main results use the well-posedness assumptions in Assumption 2 below.
Assumption 2. We require the systems (PDV) and (PDVθ _\theta θ ) to be well posed in the following sense.
2.1 The system (PDV) has a global continuous solution ( X , Y ) (X,Y) ( X , Y ) , that is unique in law.
2.2 For every bounded stopping time θ \theta θ , ( P D V ~ θ ) (\widetilde{\mathrm{PDV}}_\theta) ( PDV θ ) has a continuous solution ( X ~ θ , Y ~ θ ) (\widetilde{X}^{\theta},\widetilde{Y}^{\theta}) ( X θ , Y θ ) up to explosion, that is unique in law. In particular, for every n > 0 n>0 n > 0 , with τ ( θ , n ) X ~ θ : = inf { u ≥ 0 : ∣ X ~ θ + u θ ∣ ≥ n } \tau^{\widetilde{X}^{\theta}}_{(\theta,n)}:=\inf\{u\geq0:|\widetilde{X}^{\theta}_{\theta+u}|\geq n\} τ ( θ , n ) X θ := inf { u ≥ 0 : ∣ X θ + u θ ∣ ≥ n } , the conditional law given F θ \mathcal{F}_{\theta} F θ of ( X ~ t ∧ ( θ + τ ( θ , n ) X ~ θ ) θ , Y ~ t ∧ ( θ + τ ( θ , n ) X ~ θ ) θ ) t ≥ 0 (\widetilde{X}^{\theta}_{t\wedge(\theta+\tau^{\widetilde{X}^{\theta}}_{(\theta,n)})},\widetilde{Y}^{\theta}_{t\wedge(\theta+\tau^{\widetilde{X}^{\theta}}_{(\theta,n)})})_{t\geq0} ( X t ∧ ( θ + τ ( θ , n ) X θ ) θ , Y t ∧ ( θ + τ ( θ , n ) X θ ) θ ) t ≥ 0 is the same for all such solutions.
We evaluate a solution of ( P D V ~ θ ) (\widetilde{\mathrm{PDV}}_\theta) ( PDV θ ) only before its explosion time θ + lim n → ∞ τ ( θ , n ) X ~ θ \theta+\lim_{n\to\infty}\tau^{\widetilde{X}^{\theta}}_{(\theta,n)} θ + lim n → ∞ τ ( θ , n ) X θ . A stopped integrand is defined to be zero after its stopping time, without evaluating the solution there.
Remark 2.2. The weak well-posedness requirement is stated in Assumption 2. Under Assumption 1 and for the continuous initial curves G 1 , G 2 G_1,G_2 G 1 , G 2 in the framework, a sufficient condition for well-posedness of ( P D V ) (\mathrm{PDV}) ( PDV ) is that F 1 , F 2 F_1,F_2 F 1 , F 2 are locally Lipschitz and Γ 1 , Γ 2 \Gamma_1,\Gamma_2 Γ 1 , Γ 2 are locally Lipschitz in space, locally uniformly in time. For ( P D V ~ θ ) (\widetilde{\mathrm{PDV}}_\theta) ( PDV θ ) , it is sufficient to assume in addition that the product Γ 1 σ \Gamma_1\sigma Γ 1 σ , which appears in the additional drift, is locally Lipschitz in space, locally uniformly in time. The kernel increment bounds in Assumption 1.1 and polynomial growth bounds in Assumption 1.2 then give global well-posedness of ( P D V ) (\mathrm{PDV}) ( PDV ) and local well-posedness of ( P D V ~ θ ) (\widetilde{\mathrm{PDV}}_\theta) ( PDV θ ) up to explosion, including the conditional uniqueness in law in Assumption 2.2.
This criterion applies to Volterra–Stein–Stein with σ ( t , x ) = x \sigma(t,x)=x σ ( t , x ) = x or σ ( t , x ) = ∣ x ∣ \sigma(t,x)=|x| σ ( t , x ) = ∣ x ∣ , lognormal stochastic volatility, the Quintic model, and Volterra–Bergomi when σ \sigma σ is locally Lipschitz in space, locally uniformly in time. It also covers Quadratic Volterra Heston, since V \sqrt{V} V is smooth for a 0 > 0 a_0>0 a 0 > 0 .
Weak well-posedness can also hold when the coefficients are not locally Lipschitz on the whole state space. The following models can be treated by localization or by weak existence and uniqueness results.
PDV models. For the following models, we have K 2 , Γ 2 ≥ 0 K_2,\Gamma_2\geq0 K 2 , Γ 2 ≥ 0 , so assuming G 2 ( t ) > 0 G_2(t)>0 G 2 ( t ) > 0 for every t ≥ 0 t\geq0 t ≥ 0 gives Y 2 , t ≥ G 2 ( t ) > 0 Y_{2,t}\geq G_2(t)>0 Y 2 , t ≥ G 2 ( t ) > 0 . All lower bounds stated here also hold for the corresponding transformed processes ( P D V ~ θ ) (\widetilde{\mathrm{PDV}}_\theta) ( PDV θ ) before explosion. When F 2 ( w ) = β 2 w + F_2(w)=\beta_2\sqrt{w^+} F 2 ( w ) = β 2 w + , this prevents its singularity at zero from being reached.
Guyon–Lekeufack and specification (A). Here F 1 F_1 F 1 is locally Lipschitz. For Guyon–Lekeufack, the above localization is used in [31 , 6 ] under their respective kernel and initial-data conditions.
Specification (B). If β 0 , 1 + β 1 , 1 y 0 > β 0 , 2 \beta_{0,1}+\beta_{1,1}y_0>\beta_{0,2} β 0 , 1 + β 1 , 1 y 0 > β 0 , 2 , the affine branch dominates near y 0 y_0 y 0 , where F 1 ( y ) = β 1 , 1 y F_1(y)=\beta_{1,1}y F 1 ( y ) = β 1 , 1 y . Since both branches are locally Lipschitz away from y 0 y_0 y 0 , F 1 F_1 F 1 is locally Lipschitz on R \mathbb{R} R .
Polynomial PDV. If β 0 + F 1 ( y ) ≥ 0 \beta_0+F_1(y)\geq0 β 0 + F 1 ( y ) ≥ 0 for every y ∈ R y\in\mathbb{R} y ∈ R , then we have
∀ t ≥ 0 , X t = β 0 + F 1 ( Y 1 , t ) + β 2 Y 2 , t ≥ β 2 G 2 ( t ) > 0. \forall t\geq0,\qquad X_t=\beta_0+F_1(Y_{1,t})+\beta_2Y_{2,t}\geq\beta_2G_2(t)>0. ∀ t ≥ 0 , X t = β 0 + F 1 ( Y 1 , t ) + β 2 Y 2 , t ≥ β 2 G 2 ( t ) > 0.
Thus the singularity of Γ 1 ( x ) = x + \Gamma_1(x)=\sqrt{x^+} Γ 1 ( x ) = x + at zero is also avoided. In each case, the stated conditions recover the locally Lipschitz setting by localization.
Rough and Volterra–Heston models. Here Γ 1 ( x ) = x + \Gamma_1(x)=\sqrt{x^+} Γ 1 ( x ) = x + is not locally Lipschitz at zero. Existence and uniqueness in law are provided by the affine Volterra results of Abi Jaber, Larsson and Pulido [2 ] , under their kernel and initial-data conditions.
Martingality
Notation 1. The sign s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } specifies whether large positive or large negative values of Y 1 Y_1 Y 1 are considered. The sign ε s \varepsilon_s ε s is the sign of F 1 ( y ) F_1(y) F 1 ( y ) for sufficiently large y y y , as specified in Assumption 1.2. The sign s s s is chosen separately in each assumption or result where it is used. When ρ ≠ 0 \rho\neq0 ρ = 0 , sgn ( ρ ) \operatorname{sgn}(\rho) sgn ( ρ ) denotes the sign of ρ \rho ρ .
We first explain the sign conditions in Assumptions 3 and 5.1 below. For ρ ≠ 0 \rho\ne0 ρ = 0 and s ∈ { − 1 , 1 } s \in\{-1,1\} s ∈ { − 1 , 1 } , we have
s Y ~ 1 , t θ = s G 1 ( t ) + s ∫ 0 t K 1 ( t − u ) Γ 1 ( u , X ~ u θ ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } K 1 ( t − u ) Γ 1 ( u , X ~ u θ ) σ ( u , X ~ u θ ) d u . (19) s\widetilde{Y}^{\theta}_{1,t}=sG_{1}(t)+s\int_{0}^{t}K_{1}(t-u)\Gamma_{1}(u,\widetilde{X}^{\theta}_{u})\,\mathrm{d}W_{u}
+s\rho\int_{0}^{t}\mathbf{1}_{\{u\ge\theta\}}K_{1}(t-u)\Gamma_{1}(u,\widetilde{X}^{\theta}_{u})\sigma(u,\widetilde{X}^{\theta}_{u})\,\mathrm{d}u. \tag*{(19)} s Y 1 , t θ = s G 1 ( t ) + s ∫ 0 t K 1 ( t − u ) Γ 1 ( u , X u θ ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } K 1 ( t − u ) Γ 1 ( u , X u θ ) σ ( u , X u θ ) d u . (19)
To fix ideas, we first discuss the case where s ρ Γ 1 σ ≥ 0 s\rho\Gamma_{1}\sigma\ge0 s ρ Γ 1 σ ≥ 0 .
Case 1: Potentially destabilizing drift. When s ρ Γ 1 σ ≥ 0 s\rho\Gamma_{1}\sigma\ge0 s ρ Γ 1 σ ≥ 0 , the Girsanov drift tends to increase s Y ~ 1 θ s\widetilde{Y}^{\theta}_{1} s Y 1 θ and push it towards large positive values. By Assumption 1.2, there exists z ∗ > 0 z^{*}>0 z ∗ > 0 such that
∀ z ≥ z ∗ , ε s F 1 ( s z ) ≥ 0. \forall z\ge z^{*},\qquad\varepsilon_{s}F_{1}(sz)\ge0. ∀ z ≥ z ∗ , ε s F 1 ( sz ) ≥ 0.
If s Y ~ 1 , t θ ≥ z ∗ s\widetilde{Y}^{\theta}_{1,t}\ge z^{*} s Y 1 , t θ ≥ z ∗ , then F 1 ( Y ~ 1 , t θ ) = F 1 ( s ( s Y ~ 1 , t θ ) ) F_{1}(\widetilde{Y}^{\theta}_{1,t})=F_{1}(s(s\widetilde{Y}^{\theta}_{1,t})) F 1 ( Y 1 , t θ ) = F 1 ( s ( s Y 1 , t θ )) satisfies ε s F 1 ( Y ~ 1 , t θ ) ≥ 0 \varepsilon_{s}F_{1}(\widetilde{Y}^{\theta}_{1,t})\ge0 ε s F 1 ( Y 1 , t θ ) ≥ 0 , and
ε s X ~ t θ = ε s β ( t ) + ∣ F 1 ( Y ~ 1 , t θ ) ∣ + ε s F 2 ( Y ~ 2 , t θ ) . (20) \varepsilon_{s}\widetilde{X}^{\theta}_{t}=\varepsilon_{s}\beta(t)+\lvert F_{1}(\widetilde{Y}^{\theta}_{1,t})\rvert+\varepsilon_{s}F_{2}(\widetilde{Y}^{\theta}_{2,t}). \tag*{(20)} ε s X t θ = ε s β ( t ) + ∣ F 1 ( Y 1 , t θ )∣ + ε s F 2 ( Y 2 , t θ ) . (20)
Heuristically, growth of s Y ~ 1 θ s\widetilde{Y}^{\theta}_{1} s Y 1 θ therefore tends to increase ε s X ~ θ \varepsilon_{s}\widetilde{X}^{\theta} ε s X θ and push it towards large positive values through (20) , once the contribution of F 2 F_{2} F 2 is controlled. If s ρ Γ 1 σ s\rho\Gamma_{1}\sigma s ρ Γ 1 σ grows sufficiently in the direction, ε s x → + ∞ \varepsilon_{s}x\to+\infty ε s x → + ∞ , in the sense of (2.8, the additional drift in (19) can in turn push s Y ~ 1 θ s\widetilde{Y}^{\theta}_{1} s Y 1 θ further upward, in the direction s Y ~ 1 θ → + ∞ s\widetilde{Y}^{\theta}_{1}\to+\infty s Y 1 θ → + ∞ . The explosion mechanism is summarized in Figure 1 . The exponent χ \chi χ is defined in (23) of Assumption 5.1 below.
Figure 1. Positive feedback loop leading to explosion of (P D V ~ θ \widetilde{PDV}_{\theta} P D V θ ).
For a sufficiently short time interval [ θ , t ] [\theta,t] [ θ , t ] , the growth bounds on F 2 , Γ 2 F_{2},\Gamma_{2} F 2 , Γ 2 in Assumption 1.2, the condition γ 2 f 2 ≤ 1 \gamma_{2}f_{2}\le1 γ 2 f 2 ≤ 1 , and the regularity of K 2 K_{2} K 2 in Assumption 1.1 give, before explosion,
sup u ∈ [ θ , t ] ∣ F 2 ( Y ~ 2 , u ) ∣ ≤ B + 1 4 sup u ∈ [ 0 , t ] ∣ X ~ u ∣ , \sup_{u\in[\theta,t]}\lvert F_{2}(\widetilde{Y}_{2,u})\rvert\le B+\frac{1}{4}\sup_{u\in[0,t]}\lvert\widetilde{X}_{u}\rvert, u ∈ [ θ , t ] sup ∣ F 2 ( Y 2 , u )∣ ≤ B + 4 1 u ∈ [ 0 , t ] sup ∣ X u ∣ ,
as shown in (49) in the proof of explosion below. The following bounds hold when s Y ~ 1 s\widetilde{Y}_{1} s Y 1 remains bounded from below up to explosion. The contribution of F 1 ( Y ~ 1 ) F_{1}(\widetilde{Y}_{1}) F 1 ( Y 1 ) when s Y ~ 1 < 0 s\widetilde{Y}_{1}<0 s Y 1 < 0 is then bounded and can be absorbed into B B B . The identity
ε s X ~ u = ε s β ( u ) + ε s F 1 ( Y ~ 1 , u ) + ε s F 2 ( Y ~ 2 , u ) \varepsilon_s\widetilde{X}_u=\varepsilon_s\beta(u)+\varepsilon_sF_1(\widetilde{Y}_{1,u})+\varepsilon_sF_2(\widetilde{Y}_{2,u}) ε s X u = ε s β ( u ) + ε s F 1 ( Y 1 , u ) + ε s F 2 ( Y 2 , u )
and the polynomial bounds on F 1 F_1 F 1 in Assumption 1.2 give
c ( sup u ∈ [ θ , t ] ( s Y ~ 1 , u ) + ) f 1 , s − B ≤ sup u ∈ [ θ , t ] ε s X ~ u ≤ B + C ( sup u ∈ [ θ , t ] ( s Y ~ 1 , u ) + ) f 1 , s . c\left(\sup_{u\in[\theta,t]}(s\widetilde{Y}_{1,u})^+\right)^{f_{1,s}}-B\leq\sup_{u\in[\theta,t]}\varepsilon_s\widetilde{X}_u\leq B+C\left(\sup_{u\in[\theta,t]}(s\widetilde{Y}_{1,u})^+\right)^{f_{1,s}}. c ( u ∈ [ θ , t ] sup ( s Y 1 , u ) + ) f 1 , s − B ≤ u ∈ [ θ , t ] sup ε s X u ≤ B + C ( u ∈ [ θ , t ] sup ( s Y 1 , u ) + ) f 1 , s .
Thus, s Y ~ 1 s\widetilde{Y}_1 s Y 1 explodes towards + ∞ +\infty + ∞ if and only if ε s X ~ \varepsilon_s\widetilde{X} ε s X explodes towards + ∞ +\infty + ∞ , and their explosion times coincide.
Case 2: Stabilizing drift. When s ρ Γ 1 σ ≥ 0 s\rho\Gamma_1\sigma\geq0 s ρ Γ 1 σ ≥ 0 , the Girsanov drift stabilizes the negative side: if s Y ~ 1 θ s\widetilde{Y}_1^\theta s Y 1 θ is negative, the Girsanov drift pushes s Y ~ 1 θ s\widetilde{Y}_1^\theta s Y 1 θ towards zero and cannot drive it towards − ∞ -\infty − ∞ .
If s ρ Γ 1 σ ≤ 0 s\rho\Gamma_1\sigma\leq0 s ρ Γ 1 σ ≤ 0 , the same discussion applies with s s s replaced by − s -s − s : the roles of the positive and negative sides are interchanged.
Remark 2.3. Except for the Quintic and Stein–Stein models, all models studied here satisfy
∀ ( t , x ) ∈ R + × R , Γ 1 ( t , x ) σ ( t , x ) ≥ 0. \forall(t,x)\in\mathbb{R}_+\times\mathbb{R},\quad\Gamma_1(t,x)\sigma(t,x)\geq0. ∀ ( t , x ) ∈ R + × R , Γ 1 ( t , x ) σ ( t , x ) ≥ 0.
The choice s = sgn ( ρ ) s=\operatorname{sgn}(\rho) s = sgn ( ρ ) therefore gives s ρ Γ 1 σ ≥ 0 s\rho\Gamma_1\sigma\geq0 s ρ Γ 1 σ ≥ 0 . Where Γ 1 σ ≤ 0 \Gamma_1\sigma\leq0 Γ 1 σ ≤ 0 , choosing s = − sgn ( ρ ) s=-\operatorname{sgn}(\rho) s = − sgn ( ρ ) gives s ρ Γ 1 σ ≥ 0 s\rho\Gamma_1\sigma\geq0 s ρ Γ 1 σ ≥ 0 . For the Quintic model, Γ 1 σ \Gamma_1\sigma Γ 1 σ has the sign of x x x on both tails. When ρ > 0 \rho>0 ρ > 0 , both choices ( s , ε s ) = ( 1 , 1 ) (s,\varepsilon_s)=(1,1) ( s , ε s ) = ( 1 , 1 ) and ( − 1 , − 1 ) (-1,-1) ( − 1 , − 1 ) therefore give s ρ Γ 1 ( t , x ) σ ( t , x ) > 0 s\rho\Gamma_1(t,x)\sigma(t,x)>0 s ρ Γ 1 ( t , x ) σ ( t , x ) > 0 for sufficiently large s x sx s x .
Assumption 3. Suppose that Assumption 1 holds. We use the polynomial growth exponents from Assumption 1.2.
Non-accessibility of the explosive region
We assume ρ ≠ 0 \rho\ne0 ρ = 0 . There exists s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } such that, for every T > 0 T>0 T > 0 , there are constants L T ∈ R L_T\in\mathbb{R} L T ∈ R and C T > 0 C_T>0 C T > 0 such that
sup t ∈ [ 0 , T ] ε s X t ≤ L T P 0 -a.s. , \sup_{t\in[0,T]}\varepsilon_sX_t\leq L_T\qquad\mathbb{P}_0\text{-a.s.}, t ∈ [ 0 , T ] sup ε s X t ≤ L T P 0 -a.s. ,
and
∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≤ L T ⟹ − s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ 1 f 1 , − s ) . (21) \forall(t,x)\in[0,T]\times\mathbb{R},\qquad\varepsilon_sx\leq L_T\Longrightarrow-s\rho\Gamma_1(t,x)\sigma(t,x)\leq C_T\left(1+|x|^{\frac{1}{f_{1,-s}}}\right). \tag*{(21)} ∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≤ L T ⟹ − s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ f 1 , − s 1 ) . (21)
One-sided growth bound for the Girsanov drift
For every T > 0 T>0 T > 0 , there exist R T , C T ≥ 1 R_T,C_T\geq1 R T , C T ≥ 1 such that
∀ s ∈ { − 1 , 1 } , ∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≥ R T ⟹ s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ 1 f 1 , s ) . \forall s\in\{-1,1\},\ \forall(t,x)\in[0,T]\times\mathbb{R},\qquad\varepsilon_sx\geq R_T\Longrightarrow s\rho\Gamma_1(t,x)\sigma(t,x)\leq C_T\left(1+|x|^{\frac{1}{f_{1,s}}}\right). ∀ s ∈ { − 1 , 1 } , ∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≥ R T ⟹ s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ f 1 , s 1 ) .
Sufficient conditions for martingality are given in Assumptions 3.1 and 3.2. In Assumption 3.2, other Lyapunov functions could be used to control the solutions of ( P D V θ ) (\mathrm{PDV}_\theta) ( PDV θ ) . More precisely:
The upper bound on ε s X \varepsilon_sX ε s X is given in Assumption 3.1. By the stopped change-of-measure argument, for every T > 0 T>0 T > 0 , the bound ε s X ~ t θ ≤ L T \varepsilon_s\widetilde{X}^{\theta}_t\leq L_T ε s X t θ ≤ L T holds almost surely for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] before explosion. Thus, Assumption 3.1 rules out explosion along ε s X ~ t θ → + ∞ \varepsilon_s\widetilde{X}^{\theta}_t\to+\infty ε s X t θ → + ∞ , and hence along s Y ~ 1 , t θ → + ∞ s\widetilde{Y}^{\theta}_{1,t}\to+\infty s Y 1 , t θ → + ∞ . It remains to control the opposite direction − s Y ~ 1 , t θ → + ∞ -s\widetilde{Y}^{\theta}_{1,t}\to+\infty − s Y 1 , t θ → + ∞ through the one-sided growth bound (21) in Assumption 3.1 with exponent 1 f 1 , − s \frac{1}{f_{1,-s}} f 1 , − s 1 . If s ρ Γ 1 σ ≥ 0 s\rho\Gamma_1\sigma\geq0 s ρ Γ 1 σ ≥ 0 , (21) is trivially satisfied.
On each tail selected by ε s \varepsilon_s ε s , Assumption 3.2 controls s ρ Γ 1 ( t , x ) σ ( t , x ) s\rho\Gamma_1(t,x)\sigma(t,x) s ρ Γ 1 ( t , x ) σ ( t , x ) from above when ε s x \varepsilon_sx ε s x is large, namely ε s x ≥ R T \varepsilon_sx\geq R_T ε s x ≥ R T .
The bound (21) holds with s = sgn ( ρ ) s=\operatorname{sgn}(\rho) s = sgn ( ρ ) if
∀ ( t , x ) ∈ R + × R , Γ 1 ( t , x ) σ ( t , x ) ≥ 0. \forall(t,x)\in\mathbb{R}_+\times\mathbb{R},\qquad\Gamma_1(t,x)\sigma(t,x)\geq0. ∀ ( t , x ) ∈ R + × R , Γ 1 ( t , x ) σ ( t , x ) ≥ 0.
The one-sided drift bound in Assumption 3.2 is checked for the listed models in Table 4 .
Table 4. Verification of Assumption 3.2.
Notation 2. For a locally finite signed Borel measure μ \mu μ on R + \mathbb{R}_+ R + and a measurable function h h h , we define
∀ t ∈ [ 0 , + ∞ ) , ( h ∗ μ ) ( t ) = ( μ ∗ h ) ( t ) : = ∫ [ 0 , t ] h ( t − s ) μ ( d s ) , \forall t\in[0,+\infty),\qquad(h\ast\mu)(t)=(\mu\ast h)(t):=\int_{[0,t]}h(t-s)\,\mu(\mathrm{d}s), ∀ t ∈ [ 0 , + ∞ ) , ( h ∗ μ ) ( t ) = ( μ ∗ h ) ( t ) := ∫ [ 0 , t ] h ( t − s ) μ ( d s ) ,
whenever the integral is well defined. If h h h and g g g are two locally integrable functions on R + \mathbb{R}_+ R + , we write
( h ∗ g ) ( t ) : = ∫ 0 t h ( t − s ) g ( s ) d s , for a.e. t > 0. (h\ast g)(t):=\int_0^t h(t-s)g(s)\,\mathrm{d}s,\qquad\text{for a.e. }t>0. ( h ∗ g ) ( t ) := ∫ 0 t h ( t − s ) g ( s ) d s , for a.e. t > 0.
A locally finite signed Borel measure R 1 R_1 R 1 on R + \mathbb{R}_+ R + is called a resolvent of the first kind of a locally integrable kernel K 1 K_1 K 1 if
( K 1 ∗ R 1 ) ( t ) = ( R 1 ∗ K 1 ) ( t ) = 1 , for a.e. t > 0. (K_1\ast R_1)(t)=(R_1\ast K_1)(t)=1,\qquad\text{for a.e. }t>0. ( K 1 ∗ R 1 ) ( t ) = ( R 1 ∗ K 1 ) ( t ) = 1 , for a.e. t > 0.
Assumption 4. Kernel conditions. We use the following two conditions separately.
4.1 Nonnegative resolvent for the martingality criterion in Theorem 1(B.ii) below. The kernel K 1 K_1 K 1 belongs to L l o c 1 ( R + ) L^1_{\mathrm{loc}}(\mathbb{R}_+) L loc 1 ( R + ) , is nonnegative, and has a nonnegative resolvent of the first kind of the form
R 1 = a r e s δ 0 + ℓ r e s ( t ) d t , R_1=a_{\mathrm{res}}\delta_0+\ell_{\mathrm{res}}(t)\,\mathrm{d}t, R 1 = a res δ 0 + ℓ res ( t ) d t ,
where δ 0 \delta_0 δ 0 is the unit mass at zero, a r e s ≥ 0 a_{\mathrm{res}}\geq0 a res ≥ 0 , and ℓ r e s ∈ L l o c 1 ( R + ) \ell_{\mathrm{res}}\in L^1_{\mathrm{loc}}(\mathbb{R}_+) ℓ res ∈ L loc 1 ( R + ) is nonnegative and nonincreasing.
4.2 Smooth-kernel regularity for the explosion argument. The kernel K 1 K_1 K 1 has a continuous extension to [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) satisfying
K 1 ( 0 ) > 0 , K 1 ∈ C 1 ( [ 0 , + ∞ ) ) , K 1 ′ ∈ B V l o c ( [ 0 , + ∞ ) ) . K_1(0)>0,\qquad K_1\in C^1([0,+\infty)),\qquad K_1'\in BV_{\mathrm{loc}}([0,+\infty)). K 1 ( 0 ) > 0 , K 1 ∈ C 1 ([ 0 , + ∞ )) , K 1 ′ ∈ B V loc ([ 0 , + ∞ )) .
Here B V l o c BV_{\mathrm{loc}} B V loc denotes functions of locally bounded variation.
Remark 2.4. The main classes of kernels and possible choices of the regularity exponents are listed in Table 5 . We define the two-parameter Mittag–Leffler function, for α , β > 0 \alpha,\beta>0 α , β > 0 , by E α , β ( z ) : = ∑ k = 0 ∞ z k Γ ( α k + β ) E_{\alpha,\beta}(z):=\sum_{k=0}^{\infty}\frac{z^k}{\Gamma(\alpha k+\beta)} E α , β ( z ) := ∑ k = 0 ∞ Γ ( α k + β ) z k .
Theorem 1. Suppose that Assumption 2 holds. Then the following statements hold.
(A) If ρ = 0 \rho=0 ρ = 0 , then ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 is a martingale.
(B) Suppose in addition that ρ ≠ 0 \rho\ne0 ρ = 0 and that Assumption 1 holds. The following statements hold.
(B.i) Martingality without entering the explosive region. If Assumption 3.1 holds, then ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 is a martingale.
(B.ii) Martingality under a one-sided drift bound. If Assumptions 3.2 and 4.1 hold, then ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 is a martingale.
Strict local martingality
Assumption 5. Explosive branch and accessibility. Suppose that Assumption 1 holds. We use the notation for the polynomial-growth exponents in Assumption 1.2. We assume that ρ ≠ 0 \rho\ne0 ρ = 0 .
Growth and monotonicity conditions for explosion
There exist s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } , z 0 ∈ R z_0\in\mathbb{R} z 0 ∈ R , x ∗ ≥ 0 x^{*}\geq0 x ∗ ≥ 0 , χ > 0 \chi>0 χ > 0 , such that
z ⟼ ε s F 1 ( s z ) is nondecreasing on [ z 0 , + ∞ ) . (22) z\longmapsto\varepsilon_s F_1(sz)\ \text{is nondecreasing on }[z_0,+\infty). \tag*{(22)} z ⟼ ε s F 1 ( sz ) is nondecreasing on [ z 0 , + ∞ ) . (22)
For every T > 0 T>0 T > 0 , there exists c T > 0 c_T>0 c T > 0 such that
∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≥ x ∗ ⟹ s ρ Γ 1 ( t , x ) σ ( t , x ) ≥ c T ( ε s x ) χ . (23) \forall(t,x)\in[0,T]\times\mathbb{R},\qquad\varepsilon_s x\geq x^{*}\Longrightarrow s\rho\Gamma_1(t,x)\sigma(t,x)\geq c_T(\varepsilon_s x)^\chi. \tag*{(23)} ∀ ( t , x ) ∈ [ 0 , T ] × R , ε s x ≥ x ∗ ⟹ s ρ Γ 1 ( t , x ) σ ( t , x ) ≥ c T ( ε s x ) χ . (23)
We also assume
f 1 , s χ > 1. (24) f_{1,s}\chi>1. \tag*{(24)} f 1 , s χ > 1. (24)
Accessibility of the explosive region
With s s s , ε s \varepsilon_s ε s , x ∗ x^{*} x ∗ , z 0 z_0 z 0 as in Assumption 5.1, we define the stopping time
τ ∗ : = inf { u ≥ 0 : ε s X u > x ∗ , s Y 1 , u > z 0 } , \tau_{*}:=\inf\{u\geq0:\varepsilon_sX_u>x^{*},\ sY_{1,u}>z_0\}, τ ∗ := inf { u ≥ 0 : ε s X u > x ∗ , s Y 1 , u > z 0 } ,
and assume that
P 0 ( τ ∗ < ∞ ) > 0. \mathbb{P}_0(\tau_{*}<\infty)>0. P 0 ( τ ∗ < ∞ ) > 0.
By continuity of X , Y 1 X,Y_1 X , Y 1 , this condition is equivalent to
∃ t 0 ≥ 0 , P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0. \exists t_0 \ge0,\qquad\mathbb{P}_0(\varepsilon_sX_{t_0}>x^*,\ sY_{1,t_0}>z_0)>0. ∃ t 0 ≥ 0 , P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0.
We use + ∞ + u = + ∞ +\infty+u=+\infty + ∞ + u = + ∞ for u ≥ 0 u\ge0 u ≥ 0 . We define the sigma-field F ∞ \mathcal{F}_\infty F ∞ by F ∞ : = Gen ( ⋃ t ≥ 0 F t ) \mathcal{F}_\infty:=\operatorname{Gen}(\bigcup_{t\ge0}\mathcal{F}_t) F ∞ := Gen ( ⋃ t ≥ 0 F t ) . For a possibly infinite stopping time τ \tau τ , we define the sigma-field F τ \mathcal{F}_\tau F τ by
F τ : = { A ∈ F ∞ : ∀ t ∈ [ 0 , + ∞ ) , A ∩ { τ ≤ t } ∈ F t } . \mathcal{F}_\tau:=\{A\in\mathcal{F}_\infty:\ \forall t\in[0,+\infty),\ A\cap\{\tau\le t\}\in\mathcal{F}_t\}. F τ := { A ∈ F ∞ : ∀ t ∈ [ 0 , + ∞ ) , A ∩ { τ ≤ t } ∈ F t } .
Theorem 2. Suppose that Assumptions 1, 2 and 4.2 hold, and that Assumption 5.1 holds.
(i) For every t 0 ≥ 0 t_0\ge0 t 0 ≥ 0 and t > t 0 t>t_0 t > t 0 , if
P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 , \mathbb{P}_0(\varepsilon_sX_{t_0}>x^*,\ sY_{1,t_0}>z_0)>0, P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 , then we have E 0 [ S t ] < E 0 [ S t 0 ] \mathbb{E}_0[S_t]<\mathbb{E}_0[S_{t_0}] E 0 [ S t ] < E 0 [ S t 0 ] , and ( S u ) u ∈ [ t 0 , t ] (S_u)_{u\in[t_0,t]} ( S u ) u ∈ [ t 0 , t ] is a strict local martingale with respect to ( F u ) u ∈ [ t 0 , t ] (\mathcal{F}_u)_{u\in[t_0,t]} ( F u ) u ∈ [ t 0 , t ] .
(ii) If Assumption 5.2 holds, then, for every h > 0 h>0 h > 0 , we have
E 0 [ 1 { τ ∗ < ∞ } S τ ∗ + h ∣ F τ ∗ ] < S τ ∗ \mathbb{E}_0\left[\mathbf{1}_{\{\tau_*<\infty\}}S_{\tau_*+h}\mid\mathcal{F}_{\tau_*}\right]<S_{\tau_*} E 0 [ 1 { τ ∗ < ∞ } S τ ∗ + h ∣ F τ ∗ ] < S τ ∗ a.s. on { τ ∗ < ∞ } \{\tau_*<\infty\} { τ ∗ < ∞ } . For u ≥ 0 u\ge0 u ≥ 0 , the variable 1 { τ ∗ < ∞ } S τ ∗ + u \mathbf{1}_{\{\tau_*<\infty\}}S_{\tau_*+u} 1 { τ ∗ < ∞ } S τ ∗ + u is defined to be zero on { τ ∗ = ∞ } \{\tau_*=\infty\} { τ ∗ = ∞ } .
Remark 2.5. The parameters used in the applications are listed in Table 6 . In this table, we choose s = sgn ( ρ ) s=\operatorname{sgn}(\rho) s = sgn ( ρ ) . The accessibility properties used for strict local martingality are stated in Lemma 2.7 below. The Volterra–Bergomi row also assumes the growth condition stated in Table 1 .
Table 6. Verification of Assumption 5.1. Here τ ∗ \tau_* τ ∗ is defined in Assumption 5.2.
The following corollary applies to most models considered here, for which F 1 F_1 F 1 is linear, including the Volterra–Heston [18 ] , Bergomi [20 ] , Quadratic Volterra Heston+ [14 ] , and Guyon–Lekeufack [25 ] models.
Corollary 2.6 (Linear F 1 F_1 F 1 ). Suppose that Assumptions 1 and 2 hold, that ρ ≠ 0 \rho\ne0 ρ = 0 , that F 1 ( y ) = A y F_1(y) = Ay F 1 ( y ) = A y for some A ∈ R ∗ A \in\mathbb{R}^* A ∈ R ∗ , and that Γ 1 ( t , x ) σ ( t , x ) ≥ 0 \Gamma_1(t,x)\sigma(t,x) \ge0 Γ 1 ( t , x ) σ ( t , x ) ≥ 0 for every ( t , x ) ∈ R + × R (t,x) \in\mathbb{R}_+ \times\mathbb{R} ( t , x ) ∈ R + × R . Then the following statements hold for every T > 0 T > 0 T > 0 .
(i) The price process ( S t ) t ∈ [ 0 , T ] (S_t)_{t \in[0,T]} ( S t ) t ∈ [ 0 , T ] is a martingale if
{ P 0 ( ∀ t ∈ [ 0 , T ] , X t ≤ 0 ) = 1 , if ρ A > 0 , P 0 ( ∀ t ∈ [ 0 , T ] , X t ≥ 0 ) = 1 , if ρ A < 0. \begin{cases}
\mathbb{P}_0(\forall t \in[0,T],\ X_t \le0) = 1, & \text{if } \rho A > 0,\\
\mathbb{P}_0(\forall t \in[0,T],\ X_t \ge0) = 1, & \text{if } \rho A < 0.
\end{cases} { P 0 ( ∀ t ∈ [ 0 , T ] , X t ≤ 0 ) = 1 , P 0 ( ∀ t ∈ [ 0 , T ] , X t ≥ 0 ) = 1 , if ρ A > 0 , if ρ A < 0. (ii) For strict local martingality, we assume that Assumptions 4.2 and 5.1 hold with threshold x ∗ ≥ 0 x^* \ge0 x ∗ ≥ 0 and signs s = sgn ( ρ ) s = \operatorname{sgn}(\rho) s = sgn ( ρ ) , ε s = sgn ( ρ A ) \varepsilon_s = \operatorname{sgn}(\rho A) ε s = sgn ( ρ A ) . For every t 0 ∈ [ 0 , T ) t_0 \in[0,T) t 0 ∈ [ 0 , T ) and t 1 ∈ ( t 0 , T ] t_1 \in(t_0,T] t 1 ∈ ( t 0 , T ] , ( S u ) u ∈ [ t 0 , t 1 ] (S_u)_{u \in[t_0,t_1]} ( S u ) u ∈ [ t 0 , t 1 ] is a strict local martingale if
{ P 0 ( X t 0 > x ∗ ) > 0 , if ρ A > 0 , P 0 ( X t 0 < − x ∗ ) > 0 , if ρ A < 0. \begin{cases}
\mathbb{P}_0(X_{t_0} > x^*) > 0, & \text{if } \rho A > 0,\\
\mathbb{P}_0(X_{t_0} < -x^*) > 0, & \text{if } \rho A < 0.
\end{cases} { P 0 ( X t 0 > x ∗ ) > 0 , P 0 ( X t 0 < − x ∗ ) > 0 , if ρ A > 0 , if ρ A < 0.
Indeed, the martingality result follows from Theorem 1(B.i), restricted to [ 0 , T ] [0,T] [ 0 , T ] , using Assumption 3.1 with s = sgn ( ρ ) s = \operatorname{sgn}(\rho) s = sgn ( ρ ) , ε s = sgn ( ρ A ) \varepsilon_s = \operatorname{sgn}(\rho A) ε s = sgn ( ρ A ) , and L T = 0 L_T = 0 L T = 0 . For part (ii), the same signs give ε s F 1 ( s z ) = ∣ A ∣ z \varepsilon_s F_1(sz) = |A|z ε s F 1 ( sz ) = ∣ A ∣ z , so the growth condition in Assumption 1.2 holds with f 1 , s = 1 f_{1,s} = 1 f 1 , s = 1 , and the monotonicity condition in (2.7) of Assumption 5.1 holds for every z 0 ∈ R z_0 \in\mathbb{R} z 0 ∈ R . If P 0 ( ε s X t 0 > x ∗ ) > 0 \mathbb{P}_0(\varepsilon_s X_{t_0} > x^*) > 0 P 0 ( ε s X t 0 > x ∗ ) > 0 , then, for some z 0 < 0 z_0 < 0 z 0 < 0 , we have P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 \mathbb{P}_0(\varepsilon_s X_{t_0} > x^*,\ sY_{1,t_0} > z_0) > 0 P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 . The result therefore follows from Theorem 2 .
Lemma 2.7 (Accessibility of the explosive region). Suppose that the assumptions of Theorem 2 hold. The stopping time τ ∗ \tau_* τ ∗ defined in Assumption 5.2 satisfies
∀ T > 0 , P 0 ( τ ∗ < T ) > 0 (25) \forall T > 0,\qquad\mathbb{P}_0(\tau_* < T) > 0 \tag*{(25)} ∀ T > 0 , P 0 ( τ ∗ < T ) > 0 (25) in each of the following cases:
the lognormal stochastic volatility, Quintic, and Volterra–Bergomi models;
the polynomial PDV model and specification (A), with X 0 > 0 X_0 > 0 X 0 > 0 and the choices of x ∗ , z 0 x^*,z_0 x ∗ , z 0 in Table 6 ;
the Quadratic Volterra Heston+ model, with a 0 > 0 a_0 > 0 a 0 > 0 and 1 ≤ p ≤ 2 1 \le p \le2 1 ≤ p ≤ 2 .
In particular, Assumption 5.2 holds in these cases.
Proof. For the lognormal, Quintic, and Volterra–Bergomi models, Y 1 , t Y_{1,t} Y 1 , t is Gaussian with positive variance under P 0 \mathbb{P}_0 P 0 for every t > 0 t > 0 t > 0 . Since X t = β ( t ) + Y 1 , t X_t = \beta(t) + Y_{1,t} X t = β ( t ) + Y 1 , t and ε s = s \varepsilon_s = s ε s = s , we have
∀ t > 0 , P 0 ( ε s X t > x ∗ , s Y 1 , t > z 0 ) = P 0 ( s Y 1 , t > max { z 0 , x ∗ − s β ( t ) } ) > 0. \forall t > 0,\qquad\mathbb{P}_0(\varepsilon_s X_t > x^*,\ sY_{1,t} > z_0) = \mathbb{P}_0\left(sY_{1,t} > \max\{z_0,x^* - s\beta(t)\}\right) > 0. ∀ t > 0 , P 0 ( ε s X t > x ∗ , s Y 1 , t > z 0 ) = P 0 ( s Y 1 , t > max { z 0 , x ∗ − s β ( t )} ) > 0.
Taking t = T / 2 t = T/2 t = T /2 gives (25) . The proof for the three remaining models is given in Appendix C. □
For the Volterra–Guyon–Lekeufack model, the condition in Corollary 2.6 (ii) is P 0 ( σ t 0 < 0 ) > 0 \mathbb{P}_0(\sigma_{t_0} < 0) > 0 P 0 ( σ t 0 < 0 ) > 0 for some t 0 < T t_0 < T t 0 < T . Its validity depends on the kernels and the initial curves.
Girsanov reduction and preliminary results
Lemma 3.1 below relates the martingality of S S S to non-explosion of the solution of (PDVθ _\theta θ ).
Lemma 3.1. Suppose that Assumption 2 holds and that σ \sigma σ is locally bounded on R + × R \mathbb{R}_{+} \times\mathbb{R} R + × R . For a bounded stopping time θ \theta θ and T > 0 T > 0 T > 0 , we then have
E 0 [ S θ + T ∣ F θ ] = S θ ( 1 − lim n → ∞ P 0 ( τ ~ ( θ , n ) θ ≤ T ∣ F θ ) ) P 0 -a.s. , \mathbb{E}_{0}\left[S_{\theta+T} \mid\mathcal{F}_{\theta}\right] = S_{\theta}\left(1-\lim_{n\to\infty}\mathbb{P}_{0}\left(\left.\widetilde{\tau}^{\theta}_{(\theta,n)}\leq T\right|\mathcal{F}_{\theta}\right)\right)\quad\mathbb{P}_{0}\text{-a.s.}, E 0 [ S θ + T ∣ F θ ] = S θ ( 1 − n → ∞ lim P 0 ( τ ( θ , n ) θ ≤ T F θ ) ) P 0 -a.s. , where τ ~ ( θ , n ) θ \widetilde{\tau}^{\theta}_{(\theta,n)} τ ( θ , n ) θ is defined in Assumption 2.2. Moreover, ( S θ + t ) t ∈ [ 0 , T ] (S_{\theta+t})_{t\in[0,T]} ( S θ + t ) t ∈ [ 0 , T ] is a martingale under P 0 \mathbb{P}_{0} P 0 with respect to ( F θ + t ) t ∈ [ 0 , T ] (\mathcal{F}_{\theta+t})_{t\in[0,T]} ( F θ + t ) t ∈ [ 0 , T ] if and only if
lim n → ∞ P 0 ( τ ~ ( θ , n ) θ ≤ T ∣ F θ ) = 0 P 0 -a.s. \lim_{n\to\infty}\mathbb{P}_{0}\left(\left.\widetilde{\tau}^{\theta}_{(\theta,n)}\leq T\right|\mathcal{F}_{\theta}\right)=0\quad\mathbb{P}_{0}\text{-a.s.} n → ∞ lim P 0 ( τ ( θ , n ) θ ≤ T F θ ) = 0 P 0 -a.s.
Proof. We use the change-of-measure argument of [34 , 28 , 29 , 8 ] .
Step 1: Change of measure. We recall that W t ρ = ρ W t + 1 − ρ 2 W t ⊥ W^{\rho}_{t}=\rho W_{t}+\sqrt{1-\rho^{2}}W^{\perp}_{t} W t ρ = ρ W t + 1 − ρ 2 W t ⊥ . For n > 0 n>0 n > 0 , we set τ ( θ , n ) X : = inf { u > 0 : ∣ X θ + u ∣ ≥ n } \tau^{X}_{(\theta,n)}:=\inf\{u>0:|X_{\theta+u}|\geq n\} τ ( θ , n ) X := inf { u > 0 : ∣ X θ + u ∣ ≥ n } . Since θ \theta θ is bounded, we fix a deterministic T ~ < ∞ \widetilde{T}<\infty T < ∞ such that θ + T ≤ T ~ \theta+T\leq\widetilde{T} θ + T ≤ T a.s. We define Q n \mathbb{Q}_{n} Q n on F T ~ \mathcal{F}_{\widetilde{T}} F T by
d Q n d P 0 ∣ F T ~ : = S θ + ( T ∧ τ ( θ , n ) X ) S θ . \left.\frac{\mathrm{d}\mathbb{Q}_{n}}{\mathrm{d}\mathbb{P}_{0}}\right|_{\mathcal{F}_{\widetilde{T}}}:=\frac{S_{\theta+(T\wedge\tau^{X}_{(\theta,n)})}}{S_{\theta}}. d P 0 d Q n F T := S θ S θ + ( T ∧ τ ( θ , n ) X ) .
This ratio is the terminal value of the stochastic exponential with respect to W ρ W^{\rho} W ρ whose integrand is σ ( s , X s ) 1 { θ ≤ s ≤ θ + ( T ∧ τ ( θ , n ) X ) } \sigma(s,X_{s})\mathbf{1}_{\{\theta\leq s\leq\theta+(T\wedge\tau^{X}_{(\theta,n)})\}} σ ( s , X s ) 1 { θ ≤ s ≤ θ + ( T ∧ τ ( θ , n ) X )} . By local boundedness of σ \sigma σ , this integrand is bounded on [ 0 , T ~ ] [0,\widetilde{T}] [ 0 , T ] . By Novikov’s criterion, the exponential is a martingale. It equals 1 at θ \theta θ , so that
E 0 [ S θ + ( T ∧ τ ( θ , n ) X ) S θ ∣ F θ ] = 1. \mathbb{E}_{0}\left[\left.\frac{S_{\theta+(T\wedge\tau^{X}_{(\theta,n)})}}{S_{\theta}}\right|\mathcal{F}_{\theta}\right]=1. E 0 [ S θ S θ + ( T ∧ τ ( θ , n ) X ) F θ ] = 1.
Thus Q n \mathbb{Q}_{n} Q n and P 0 \mathbb{P}_{0} P 0 agree on F θ \mathcal{F}_{\theta} F θ .
Step 2: Identification with ( P D V θ ) ~ (\widetilde{\mathbb{P}\mathbb{D}\mathbb{V}_{\theta})} ( PD V θ ) . Girsanov’s theorem shows that
( W t − ρ ∫ 0 t σ ( s , X s ) 1 { θ ≤ s ≤ θ + ( T ∧ τ ( θ , n ) X ) } d s ) t ∈ [ 0 , T ~ ] \left(W_{t}-\rho\int_{0}^{t}\sigma(s,X_{s})\mathbf{1}_{\{\theta\leq s\leq\theta+(T\wedge\tau^{X}_{(\theta,n)})\}}\,\mathrm{d}s\right)_{t\in[0,\widetilde{T}]} ( W t − ρ ∫ 0 t σ ( s , X s ) 1 { θ ≤ s ≤ θ + ( T ∧ τ ( θ , n ) X )} d s ) t ∈ [ 0 , T ]
is a Brownian motion under Q n \mathbb{Q}_{n} Q n and agrees with W W W on [ 0 , θ ] [0,\theta] [ 0 , θ ] . Under Q n \mathbb{Q}_{n} Q n , ( X , Y ) (X,Y) ( X , Y ) solves ( P D V θ ) ~ (\widetilde{\mathbb{P}\mathbb{D}\mathbb{V}_{\theta})} ( PD V θ ) , driven by this Brownian motion, up to θ + ( T ∧ τ ( θ , n ) X ) \theta+(T\wedge\tau^{X}_{(\theta,n)}) θ + ( T ∧ τ ( θ , n ) X ) , with history ( X s , Y s ) s ∈ [ 0 , θ ] (X_{s},Y_{s})_{s\in[0,\theta]} ( X s , Y s ) s ∈ [ 0 , θ ] . Here L Q ( Z ∣ F θ ) \mathcal{L}_{\mathbb{Q}}(Z\mid\mathcal{F}_{\theta}) L Q ( Z ∣ F θ ) denotes the conditional law of a random element Z Z Z under the probability measure Q \mathbb{Q} Q , given F θ \mathcal{F}_{\theta} F θ . Conditional uniqueness in law in Assumption 2.2 therefore gives
L Q n ( ( X θ + ( u ∧ τ ( θ , n ) X ) , Y θ + ( u ∧ τ ( θ , n ) X ) ) u ∈ [ 0 , T ] ∣ F θ ) = L P 0 ( ( X ~ θ + ( u ∧ τ ~ ( θ , n ) θ ) θ , Y ~ θ + ( u ∧ τ ~ ( θ , n ) θ ) θ ) u ∈ [ 0 , T ] ∣ F θ ) . (26) \mathcal{L}_{\mathbb{Q}_{n}}\left(\left.(X_{\theta+(u\wedge\tau^{X}_{(\theta,n)})},Y_{\theta+(u\wedge\tau^{X}_{(\theta,n)})})_{u\in[0,T]}\right|\mathcal{F}_{\theta}\right)
=
\mathcal{L}_{\mathbb{P}_{0}}\left(\left.(\widetilde{X}^{\theta}_{\theta+(u\wedge\widetilde{\tau}^{\theta}_{(\theta,n)})},\widetilde{Y}^{\theta}_{\theta+(u\wedge\widetilde{\tau}^{\theta}_{(\theta,n)})})_{u\in[0,T]}\right|\mathcal{F}_{\theta}\right). \tag*{(26)} L Q n ( ( X θ + ( u ∧ τ ( θ , n ) X ) , Y θ + ( u ∧ τ ( θ , n ) X ) ) u ∈ [ 0 , T ] F θ ) = L P 0 ( ( X θ + ( u ∧ τ ( θ , n ) θ ) θ , Y θ + ( u ∧ τ ( θ , n ) θ ) θ ) u ∈ [ 0 , T ] F θ ) . (26)
The definition of Q n \mathbb{Q}_{n} Q n and (26) give
E 0 [ S θ + T S θ 1 { τ ( θ , n ) X > T } ∣ F θ ] = Q n ( τ ( θ , n ) X > T ∣ F θ ) = 1 − P 0 ( τ ~ ( θ , n ) θ ≤ T ∣ F θ ) . \mathbb{E}_{0}\left[\left.\frac{S_{\theta+T}}{S_{\theta}}\mathbf{1}_{\{\tau^{X}_{(\theta,n)}>T\}}\right|\mathcal{F}_{\theta}\right]
=
\mathbb{Q}_{n}\left(\left.\tau^{X}_{(\theta,n)}>T\right|\mathcal{F}_{\theta}\right)
=
1-\mathbb{P}_{0}\left(\left.\widetilde{\tau}^{\theta}_{(\theta,n)}\leq T\right|\mathcal{F}_{\theta}\right). E 0 [ S θ S θ + T 1 { τ ( θ , n ) X > T } F θ ] = Q n ( τ ( θ , n ) X > T F θ ) = 1 − P 0 ( τ ( θ , n ) θ ≤ T F θ ) .
By global existence in Assumption 2.1, we have 1 { τ ( θ , n ) X > T } ↑ n → ∞ 1 \mathbf{1}_{\{\tau^{X}_{(\theta,n)}>T\}}\mathrel{\uparrow}_{n\to\infty}1 1 { τ ( θ , n ) X > T } ↑ n → ∞ 1 a.s. Conditional monotone convergence gives
E 0 [ S θ + T S θ ∣ F θ ] = 1 − lim n → ∞ P 0 ( τ ~ ( θ , n ) θ ≤ T ∣ F θ ) . \mathbb{E}_{0}\left[\left.\frac{S_{\theta+T}}{S_{\theta}}\right|\mathcal{F}_{\theta}\right]
=
1-\lim_{n\to\infty}\mathbb{P}_{0}\left(\left.\widetilde{\tau}^{\theta}_{(\theta,n)}\leq T\right|\mathcal{F}_{\theta}\right). E 0 [ S θ S θ + T F θ ] = 1 − n → ∞ lim P 0 ( τ ( θ , n ) θ ≤ T F θ ) .
Since S S S is a nonnegative local martingale under P 0 \mathbb{P}_{0} P 0 , it is a supermartingale, and ( S θ + t ) t ∈ [ 0 , T ] (S_{\theta+t})_{t\in[0,T]} ( S θ + t ) t ∈ [ 0 , T ] is a martingale if and only if E 0 [ S θ + T ∣ F θ ] = S θ \mathbb{E}_{0}[S_{\theta+T}\mid\mathcal{F}_{\theta}]=S_{\theta} E 0 [ S θ + T ∣ F θ ] = S θ P 0 \mathbb{P}_{0} P 0 -a.s., which is equivalent to
lim n → ∞ P 0 ( τ ~ ( θ , n ) θ ≤ T ∣ F θ ) = 0 P 0 -a.s. \lim_{n\to\infty}\mathbb{P}_{0}\left(\left.\widetilde{\tau}^{\theta}_{(\theta,n)}\leq T\right|\mathcal{F}_{\theta}\right)=0\quad\mathbb{P}_{0}\text{-a.s.} n → ∞ lim P 0 ( τ ( θ , n ) θ ≤ T F θ ) = 0 P 0 -a.s.
Proof of Theorem 1(A). If ρ = 0 \rho= 0 ρ = 0 , W ⊥ W^\perp W ⊥ is independent of the factor system. For every t ≥ 0 t \ge0 t ≥ 0 , conditionally on ( X u ) u ∈ [ 0 , t ] (X_u)_{u\in[0,t]} ( X u ) u ∈ [ 0 , t ] , the stochastic exponential S t / S 0 S_t/S_0 S t / S 0 has a deterministic square-integrable integrand and therefore has expectation 1. Hence we have E 0 [ S t ] = S 0 \mathbb{E}_0[S_t] = S_0 E 0 [ S t ] = S 0 for every t ≥ 0 t \ge0 t ≥ 0 , and the nonnegative local martingale S S S is a martingale. □
Martingality: Proof of Theorem 1
Throughout this section, T > 0 T > 0 T > 0 is fixed and θ = 0 \theta= 0 θ = 0 . We write X ~ \widetilde{X} X , Y ~ 1 \widetilde{Y}_1 Y 1 , and Y ~ 2 \widetilde{Y}_2 Y 2 instead of X ~ 0 \widetilde{X}^0 X 0 , Y ~ 1 0 \widetilde{Y}_1^0 Y 1 0 , and Y ~ 2 0 \widetilde{Y}_2^0 Y 2 0 , respectively. We set
τ n : = inf { t ≥ 0 : ∣ X ~ t ∣ ≥ n } . \tau_n := \inf\{t \ge0 : |\widetilde{X}_t| \ge n\}. τ n := inf { t ≥ 0 : ∣ X t ∣ ≥ n } .
The constants C C C , with or without subscripts, may depend on the initial curves and change from line to line. They are independent of n n n .
Lemma 4.1. Suppose that Assumptions 1 and 2 hold. The process defined by
∀ t ∈ [ 0 , T ] , U 1 , t n : = ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X ~ s ) 1 { s ≤ τ n } d W s , \forall t \in[0,T],\qquad U^n_{1,t} := \int_0^t K_1(t-s)\Gamma_1(s,\widetilde{X}_s)\mathbf{1}_{\{s\le\tau_n\}}\,\mathrm{d}W_s, ∀ t ∈ [ 0 , T ] , U 1 , t n := ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X s ) 1 { s ≤ τ n } d W s , admits a continuous modification. We use this continuous modification throughout the remainder of the paper. For some constant C T C_T C T , we assume that, for every n ∈ N ∗ n \in\mathbb{N}^* n ∈ N ∗ , a.s.,
∀ t ∈ [ 0 , T ] , sup u ∈ [ 0 , t ] ∣ X ~ u ∧ τ n ∣ ≤ C T ( 1 + sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ f 1 + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 ) . (27) \forall t \in[0,T],\qquad\sup_{u\in[0,t]}|\widetilde{X}_{u\wedge\tau_n}| \le C_T\left(1+\sup_{u\in[0,t]}|U^n_{1,u}|^{f_1}+\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{f_2}\right). \tag*{(27)} ∀ t ∈ [ 0 , T ] , u ∈ [ 0 , t ] sup ∣ X u ∧ τ n ∣ ≤ C T ( 1 + u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ f 1 + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 2 ) . (27) Then, for every r > 0 r > 0 r > 0 , we have
sup n ∈ N ∗ E 0 [ sup t ∈ [ 0 , T ] ∣ X ~ t ∧ τ n ∣ r ] ≤ C T , r , (28) \sup_{n\in\mathbb{N}^*}\mathbb{E}_0\left[\sup_{t\in[0,T]}|\widetilde{X}_{t\wedge\tau_n}|^r\right]\le C_{T,r}, \tag*{(28)} n ∈ N ∗ sup E 0 [ t ∈ [ 0 , T ] sup ∣ X t ∧ τ n ∣ r ] ≤ C T , r , (28) and ( S t ) t ∈ [ 0 , T ] (S_t)_{t\in[0,T]} ( S t ) t ∈ [ 0 , T ] is a martingale.
Proof. By Hölder’s inequality, we only need to prove (28) of Lemma 4.1 for r > 0 r > 0 r > 0 such that r f 1 > 2 rf_1 > 2 r f 1 > 2 , r δ 1 f 1 > 2 r\delta_1f_1 > 2 r δ 1 f 1 > 2 , and r f 2 ≥ 1 rf_2 \ge1 r f 2 ≥ 1 . We fix such an r r r and n ∈ N ∗ n \in\mathbb{N}^* n ∈ N ∗ , and set
∀ t ∈ [ 0 , T ] , A n ( t ) : = E 0 [ sup u ∈ [ 0 , t ] ∣ X ~ u ∧ τ n ∣ r ] . \forall t \in[0,T],\qquad A_n(t) := \mathbb{E}_0\left[\sup_{u\in[0,t]}|\widetilde{X}_{u\wedge\tau_n}|^r\right]. ∀ t ∈ [ 0 , T ] , A n ( t ) := E 0 [ u ∈ [ 0 , t ] sup ∣ X u ∧ τ n ∣ r ] .
The function A n A_n A n is nondecreasing. Continuity of X ~ \widetilde{X} X gives A n ( t ) ≤ ( n ∨ ∣ X ~ 0 ∣ ) r < ∞ A_n(t) \le(n\vee|\widetilde{X}_0|)^r < \infty A n ( t ) ≤ ( n ∨ ∣ X 0 ∣ ) r < ∞ .
Step 1: Stochastic convolution estimate. By the polynomial growth bound on Γ 1 \Gamma_1 Γ 1 in Assumption 1.2 and γ 1 f 1 ≤ 1 \gamma_1f_1 \le1 γ 1 f 1 ≤ 1 , we have
∀ s ∈ [ 0 , T ] , E 0 [ ∣ Γ 1 ( s , X ~ s ) ∣ r f 1 1 { s ≤ τ n } ] ≤ C T , r E 0 [ ( 1 + ∣ X ~ s ∣ r ) 1 { s ≤ τ n } ] ≤ C T , r ( 1 + A n ( s ) ) . \forall s \in[0,T],\qquad\mathbb{E}_0\left[|\Gamma_1(s,\widetilde{X}_s)|^{rf_1}\mathbf{1}_{\{s\le\tau_n\}}\right]\le C_{T,r}\mathbb{E}_0\left[(1+|\widetilde{X}_s|^r)\mathbf{1}_{\{s\le\tau_n\}}\right]\le C_{T,r}(1+A_n(s)). ∀ s ∈ [ 0 , T ] , E 0 [ ∣ Γ 1 ( s , X s ) ∣ r f 1 1 { s ≤ τ n } ] ≤ C T , r E 0 [ ( 1 + ∣ X s ∣ r ) 1 { s ≤ τ n } ] ≤ C T , r ( 1 + A n ( s )) .
For u , v ∈ [ 0 , T ] u,v \in[0,T] u , v ∈ [ 0 , T ] with u < v u<v u < v , the BDG inequality gives
E 0 [ ∣ U 1 , v n − U 1 , u n ∣ r f 1 ] ≤ C T , r E 0 [ ( ∫ 0 u ∣ K 1 ( v − s ) − K 1 ( u − s ) ∣ 2 ∣ Γ 1 ( s , X ~ s ) ∣ 2 1 { s ≤ τ n } d s + ∫ u v K 1 ( v − s ) 2 ∣ Γ 1 ( s , X ~ s ) ∣ 2 1 { s ≤ τ n } d s ) r f 1 2 ] ≤ C T , r ( 1 + A n ( v ) ) ( ∫ 0 u ∣ K 1 ( v − s ) − K 1 ( u − s ) ∣ 2 d s + ∫ u v K 1 ( v − s ) 2 d s ) r f 1 2 ≤ C T , r ∣ v − u ∣ r δ 1 f 1 ( 1 + A n ( v ) ) . (29) \begin{aligned}
\mathbb{E}_0\left[|U^n_{1,v}-U^n_{1,u}|^{rf_1}\right]
&\le C_{T,r}\mathbb{E}_0\left[\left(\int_0^u|K_1(v-s)-K_1(u-s)|^2|\Gamma_1(s,\widetilde{X}_s)|^2\mathbf{1}_{\{s\le\tau_n\}}\,\mathrm{d}s\right.\right.\\
&\qquad\left.\left.+\int_u^vK_1(v-s)^2|\Gamma_1(s,\widetilde{X}_s)|^2\mathbf{1}_{\{s\le\tau_n\}}\,\mathrm{d}s\right)^{\frac{rf_1}{2}}\right]\\
&\le C_{T,r}(1+A_n(v))\left(\int_0^u|K_1(v-s)-K_1(u-s)|^2\,\mathrm{d}s+\int_u^vK_1(v-s)^2\,\mathrm{d}s\right)^{\frac{rf_1}{2}}\\
&\le C_{T,r}|v-u|^{r\delta_1f_1}(1+A_n(v)). \tag*{(29)}
\end{aligned} E 0 [ ∣ U 1 , v n − U 1 , u n ∣ r f 1 ] ≤ C T , r E 0 [ ( ∫ 0 u ∣ K 1 ( v − s ) − K 1 ( u − s ) ∣ 2 ∣ Γ 1 ( s , X s ) ∣ 2 1 { s ≤ τ n } d s + ∫ u v K 1 ( v − s ) 2 ∣ Γ 1 ( s , X s ) ∣ 2 1 { s ≤ τ n } d s ) 2 r f 1 ≤ C T , r ( 1 + A n ( v )) ( ∫ 0 u ∣ K 1 ( v − s ) − K 1 ( u − s ) ∣ 2 d s + ∫ u v K 1 ( v − s ) 2 d s ) 2 r f 1 ≤ C T , r ∣ v − u ∣ r δ 1 f 1 ( 1 + A n ( v )) . (29)
The second inequality uses weighted Jensen’s inequality and monotonicity of A n A_n A n . The last inequality follows from the increment bound on K 1 K_1 K 1 in (16) of Assumption 1. Since A n ( T ) < ∞ A_n(T)<\infty A n ( T ) < ∞ , the increment estimate (29) , with r δ 1 f 1 > 2 > 1 r\delta_1f_1>2>1 r δ 1 f 1 > 2 > 1 , gives a continuous modification of U 1 U^1 U 1 by Kolmogorov’s continuity theorem. (This ensures that sup t ∈ [ 0 , T ] ∣ U 1 , t n ∣ \sup_{t\in[0,T]}|U^n_{1,t}| sup t ∈ [ 0 , T ] ∣ U 1 , t n ∣ is well defined.) Since U 1 , 0 n = 0 U^n_{1,0}=0 U 1 , 0 n = 0 , Lemma A.1 below with exponent r f 1 > 2 rf_1>2 r f 1 > 2 gives the first inequality below. The second follows from (29) by symmetry and Tonelli’s theorem.
∀ t ∈ [ 0 , T ] , E 0 [ sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ r f 1 ] ≤ C T , r t ∫ 0 t ∫ 0 t E 0 [ ∣ U 1 , v n − U 1 , u n ∣ r f 1 ] ∣ v − u ∣ 3 d u d v ≤ C T , r t ∫ 0 t ( 1 + A n ( v ) ) ∫ 0 v ( v − u ) r δ 1 f 1 − 3 d u d v = C T , r t r δ 1 f 1 − 2 ∫ 0 t ( 1 + A n ( v ) ) v r δ 1 f 1 − 2 d v ≤ C T , r ( 1 + ∫ 0 t A n ( v ) d v ) . (30) \begin{aligned}
\forall t\in[0,T],\qquad\mathbb{E}_0\left[\sup_{u\in[0,t]}|U^n_{1,u}|^{rf_1}\right]
&\le C_{T,r}t\int_0^t\int_0^t\frac{\mathbb{E}_0[|U^n_{1,v}-U^n_{1,u}|^{rf_1}]}{|v-u|^3}\,du\,dv\\
&\le C_{T,r}t\int_0^t(1+A_n(v))\int_0^v(v-u)^{r\delta_1f_1-3}\,du\,dv\\
&=\frac{C_{T,r}t}{r\delta_1f_1-2}\int_0^t(1+A_n(v))v^{r\delta_1f_1-2}\,dv\\
&\le C_{T,r}\left(1+\int_0^t A_n(v)\,dv\right). \tag*{(30)}
\end{aligned} ∀ t ∈ [ 0 , T ] , E 0 [ u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ r f 1 ] ≤ C T , r t ∫ 0 t ∫ 0 t ∣ v − u ∣ 3 E 0 [ ∣ U 1 , v n − U 1 , u n ∣ r f 1 ] d u d v ≤ C T , r t ∫ 0 t ( 1 + A n ( v )) ∫ 0 v ( v − u ) r δ 1 f 1 − 3 d u d v = r δ 1 f 1 − 2 C T , r t ∫ 0 t ( 1 + A n ( v )) v r δ 1 f 1 − 2 d v ≤ C T , r ( 1 + ∫ 0 t A n ( v ) d v ) . (30)
The last inequality follows from r δ 1 f 1 > 2 r\delta_1f_1>2 r δ 1 f 1 > 2 and 0 ≤ v ≤ t ≤ T 0\le v\le t\le T 0 ≤ v ≤ t ≤ T .
Step 2: Stopped moments of X ~ \widetilde{X} X . For a nonnegative nondecreasing function φ \varphi φ , nonnegativity of K 2 K_2 K 2 gives
∀ 0 ≤ u ≤ t ≤ T , ∫ 0 u K 2 ( u − s ) φ ( s ) d s = ∫ 0 u K 2 ( v ) φ ( u − v ) d v ≤ ∫ 0 u K 2 ( v ) φ ( t − v ) d v ≤ ∫ 0 t K 2 ( v ) φ ( t − v ) d v . \forall0\le u\le t\le T,\qquad
\int_0^u K_2(u-s)\varphi(s)\,ds
=\int_0^u K_2(v)\varphi(u-v)\,dv
\le\int_0^u K_2(v)\varphi(t-v)\,dv
\le\int_0^t K_2(v)\varphi(t-v)\,dv. ∀0 ≤ u ≤ t ≤ T , ∫ 0 u K 2 ( u − s ) φ ( s ) d s = ∫ 0 u K 2 ( v ) φ ( u − v ) d v ≤ ∫ 0 u K 2 ( v ) φ ( t − v ) d v ≤ ∫ 0 t K 2 ( v ) φ ( t − v ) d v .
Thus the convolution is nondecreasing. For a continuous function g g g on [ 0 , T ] [0,T] [ 0 , T ] , we write ∥ g ∥ ∞ , [ 0 , T ] : = sup t ∈ [ 0 , T ] ∣ g ( t ) ∣ \|g\|_{\infty,[0,T]}:=\sup_{t\in[0,T]}|g(t)| ∥ g ∥ ∞ , [ 0 , T ] := sup t ∈ [ 0 , T ] ∣ g ( t ) ∣ . The polynomial growth bound on Γ 2 \Gamma_2 Γ 2 in Assumption 1.2 gives
∀ t ∈ [ 0 , T ] , sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ ≤ ∥ G 2 ∥ ∞ , [ 0 , T ] + C T ∫ 0 t K 2 ( t − s ) ( 1 + sup v ∈ [ 0 , s ] ∣ X ~ v ∧ τ n ∣ γ 2 ) d s . \forall t\in[0,T],\qquad
\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|
\le\|G_2\|_{\infty,[0,T]}+C_T\int_0^t K_2(t-s)\left(1+\sup_{v\in[0,s]}|\widetilde{X}_{v\wedge\tau_n}|^{\gamma_2}\right)\,ds. ∀ t ∈ [ 0 , T ] , u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ ≤ ∥ G 2 ∥ ∞ , [ 0 , T ] + C T ∫ 0 t K 2 ( t − s ) ( 1 + v ∈ [ 0 , s ] sup ∣ X v ∧ τ n ∣ γ 2 ) d s .
Weighted Jensen’s inequality with exponent r f 2 ≥ 1 rf_2\ge1 r f 2 ≥ 1 , K 2 ∈ L 1 ( [ 0 , T ] ) K_2\in L^1([0,T]) K 2 ∈ L 1 ([ 0 , T ]) , and γ 2 f 2 ≤ 1 \gamma_2f_2\le1 γ 2 f 2 ≤ 1 give
∀ t ∈ [ 0 , T ] , E 0 [ sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ r f 2 ] ≤ C T , r ( 1 + ∫ 0 t K 2 ( t − s ) E 0 [ sup v ∈ [ 0 , s ] ∣ X ~ v ∧ τ n ∣ r γ 2 f 2 ] d s ) ≤ C T , r ( 1 + ∫ 0 t K 2 ( t − s ) A n ( s ) d s ) . \begin{aligned}
\forall t\in[0,T],\qquad
\mathbb{E}_0\left[\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{rf_2}\right]
&\le C_{T,r}\left(1+\int_0^t K_2(t-s)\mathbb{E}_0\left[\sup_{v\in[0,s]}|\widetilde{X}_{v\wedge\tau_n}|^{r\gamma_2f_2}\right]\,ds\right)\\
&\le C_{T,r}\left(1+\int_0^t K_2(t-s)A_n(s)\,ds\right).
\end{aligned} ∀ t ∈ [ 0 , T ] , E 0 [ u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ r f 2 ] ≤ C T , r ( 1 + ∫ 0 t K 2 ( t − s ) E 0 [ v ∈ [ 0 , s ] sup ∣ X v ∧ τ n ∣ r γ 2 f 2 ] d s ) ≤ C T , r ( 1 + ∫ 0 t K 2 ( t − s ) A n ( s ) d s ) .
Combining this bound with (27) of Lemma 4.1 and (30) , we obtain
∀ t ∈ [ 0 , T ] , A n ( t ) ≤ C T , r ( 1 + ∫ 0 t [ 1 + K 2 ( t − s ) ] A n ( s ) d s ) . \forall t\in[0,T],\qquad
A_n(t)\le C_{T,r}\left(1+\int_0^t[1+K_2(t-s)]A_n(s)\,ds\right). ∀ t ∈ [ 0 , T ] , A n ( t ) ≤ C T , r ( 1 + ∫ 0 t [ 1 + K 2 ( t − s )] A n ( s ) d s ) .
The Volterra–Grönwall lemma [37 Lemma 2.2] , applied with the integrable kernel C T , r ( 1 + K 2 ) C_{T,r}(1+K_2) C T , r ( 1 + K 2 ) , proves (28) of Lemma 4.1. Continuity of X ~ \widetilde{X} X and Markov’s inequality give
P 0 ( τ n ≤ T ) ≤ n − r E 0 [ sup t ∈ [ 0 , T ] ∣ X ~ t ∧ τ n ∣ r ] ≤ C T , r n − r → n → ∞ 0. \mathbb{P}_0(\tau_n\le T)\le n^{-r}\mathbb{E}_0\left[\sup_{t\in[0,T]}|\widetilde{X}_{t\wedge\tau_n}|^r\right]\le C_{T,r}n^{-r}\xrightarrow[n\to\infty]{}0. P 0 ( τ n ≤ T ) ≤ n − r E 0 [ t ∈ [ 0 , T ] sup ∣ X t ∧ τ n ∣ r ] ≤ C T , r n − r n → ∞ 0.
Lemma 3.1 with θ = 0 \theta=0 θ = 0 proves martingality on [ 0 , T ] [0,T] [ 0 , T ] . □
Proof of Theorem 1(B.ii)
Lemma 4.2 (Volterra chain-rule). Suppose that Assumption 4.1 holds. In particular, K 1 ∈ L 1 ( [ 0 , T ] ) K_1 \in L^1([0,T]) K 1 ∈ L 1 ([ 0 , T ]) and K 1 ≥ 0 K_1 \ge0 K 1 ≥ 0 . Let the function g g g belong to L ∞ ( [ 0 , T ] ) L^\infty([0,T]) L ∞ ([ 0 , T ]) , and let the process B B B be defined by
∀ t ∈ [ 0 , T ] , B t : = ∫ 0 t K 1 ( t − s ) g s d s . \forall t \in[0,T], \qquad B_t := \int_0^t K_1(t-s)g_s\,\mathrm{d}s. ∀ t ∈ [ 0 , T ] , B t := ∫ 0 t K 1 ( t − s ) g s d s . Let the function Φ : R → R + \Phi:\mathbb{R}\to\mathbb{R}_+ Φ : R → R + belong to C 1 ( R ) C^1(\mathbb{R}) C 1 ( R ) , be convex, and satisfy Φ ( 0 ) = 0 \Phi(0)=0 Φ ( 0 ) = 0 . Then we have
∀ t ∈ [ 0 , T ] , Φ ( B t ) ≤ ∫ 0 t K 1 ( t − s ) Φ ′ ( B s ) g s d s . (31) \forall t \in[0,T], \qquad\Phi(B_t) \le\int_0^t K_1(t-s)\Phi'(B_s)g_s\,\mathrm{d}s. \tag*{(31)} ∀ t ∈ [ 0 , T ] , Φ ( B t ) ≤ ∫ 0 t K 1 ( t − s ) Φ ′ ( B s ) g s d s . (31)
For a r e s = 0 a_{\mathrm{res}}=0 a res = 0 , ( k , ℓ ) = ( ℓ r e s , K 1 ) (k,\ell)=(\ell_{\mathrm{res}},K_1) ( k , ℓ ) = ( ℓ res , K 1 ) satisfies condition ( P C ) (\mathcal{PC}) ( P C ) . The convexity inequality and regularization argument of Vergara–Zacher [36, Section 2, Lemma 2.1; Section 3, proof of Lemma 3.2], with u = B u=B u = B , u 0 = 0 u_0=0 u 0 = 0 , and H = Φ H=\Phi H = Φ , give (31) , since ∂ t ( ℓ r e s ∗ B ) = g \partial_t(\ell_{\mathrm{res}}*B)=g ∂ t ( ℓ res ∗ B ) = g . The extension to a r e s > 0 a_{\mathrm{res}}>0 a res > 0 follows immediately by the ordinary chain rule for the atomic contribution: it is the ( P C ′ ) (\mathcal{PC}') ( P C ′ ) extended setting described in [36 Section 1] .
We set
∀ t ∈ [ 0 , T ] , Z t : = Y ~ 1 , t − G 1 ( t ) . \forall t \in[0,T], \qquad Z_t := \widetilde{Y}_{1,t}-G_1(t). ∀ t ∈ [ 0 , T ] , Z t := Y 1 , t − G 1 ( t ) .
Under the resolvent condition in Assumption 4.1, Z Z Z need not be a semimartingale. Thus Itô’s formula cannot be used directly in Theorem 1(B.ii). For n ∈ N ∗ n\in\mathbb{N}^* n ∈ N ∗ , we write
∀ t ∈ [ 0 , T ∧ τ n ] , Z t = U 1 , t n + B t , B t : = ρ ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X ~ s ) σ ( s , X ~ s ) d s . \forall t \in[0,T\wedge\tau_n], \qquad Z_t=U^n_{1,t}+B_t,\qquad B_t:=\rho\int_0^t K_1(t-s)\Gamma_1(s,\widetilde{X}_s)\sigma(s,\widetilde{X}_s)\,\mathrm{d}s. ∀ t ∈ [ 0 , T ∧ τ n ] , Z t = U 1 , t n + B t , B t := ρ ∫ 0 t K 1 ( t − s ) Γ 1 ( s , X s ) σ ( s , X s ) d s .
We estimate U 1 n U^n_1 U 1 n by (30) .
Proof of Theorem 1(B.ii). For t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , y , w ∈ R y,w\in\mathbb{R} y , w ∈ R , we set x = β ( t ) + F 1 ( y ) + F 2 ( w ) x=\beta(t)+F_1(y)+F_2(w) x = β ( t ) + F 1 ( y ) + F 2 ( w ) . By Assumption 1.2 and the boundedness of β \beta β on [ 0 , T ] [0,T] [ 0 , T ] , there exist 0 < c ≤ C 0<c\le C 0 < c ≤ C such that, for a sufficiently large C T ≥ 1 C_T\ge1 C T ≥ 1 and each s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } , the conditions
s y > 0 , ∣ y ∣ f 1 , s ≥ C T ( 1 + ∣ w ∣ f 2 ) sy>0,\qquad|y|^{f_{1,s}}\ge C_T(1+|w|^{f_2}) sy > 0 , ∣ y ∣ f 1 , s ≥ C T ( 1 + ∣ w ∣ f 2 )
imply ∣ y ∣ ≥ 1 |y|\ge1 ∣ y ∣ ≥ 1 and
c ∣ y ∣ f 1 , s ≤ ε s F 1 ( y ) ≤ C ∣ y ∣ f 1 , s , c|y|^{f_{1,s}}\le\varepsilon_sF_1(y)\le C|y|^{f_{1,s}}, c ∣ y ∣ f 1 , s ≤ ε s F 1 ( y ) ≤ C ∣ y ∣ f 1 , s ,
∣ β ( t ) ∣ + ∣ F 2 ( w ) ∣ ≤ C ( 1 + ∣ w ∣ f 2 ) ≤ c 2 ∣ y ∣ f 1 , s . |\beta(t)|+|F_2(w)|\le C(1+|w|^{f_2})\le\frac{c}{2}|y|^{f_{1,s}}. ∣ β ( t ) ∣ + ∣ F 2 ( w ) ∣ ≤ C ( 1 + ∣ w ∣ f 2 ) ≤ 2 c ∣ y ∣ f 1 , s .
With R T R_T R T as in Assumption 3.2, increasing C T C_T C T if necessary yields
ε s x ≥ ε s F 1 ( y ) − ∣ β ( t ) ∣ − ∣ F 2 ( w ) ∣ ≥ c 2 ∣ y ∣ f 1 , s ≥ R T , \varepsilon_sx\ge\varepsilon_sF_1(y)-|\beta(t)|-|F_2(w)|\ge\frac{c}{2}|y|^{f_{1,s}}\ge R_T, ε s x ≥ ε s F 1 ( y ) − ∣ β ( t ) ∣ − ∣ F 2 ( w ) ∣ ≥ 2 c ∣ y ∣ f 1 , s ≥ R T ,
ε s x ≤ ε s F 1 ( y ) + ∣ β ( t ) ∣ + ∣ F 2 ( w ) ∣ ≤ 2 C ∣ y ∣ f 1 , s . (32) \varepsilon_sx\le\varepsilon_sF_1(y)+|\beta(t)|+|F_2(w)|\le2C|y|^{f_{1,s}}. \tag*{(32)} ε s x ≤ ε s F 1 ( y ) + ∣ β ( t ) ∣ + ∣ F 2 ( w ) ∣ ≤ 2 C ∣ y ∣ f 1 , s . (32)
By (32) , we have ∣ x ∣ = ε s x ≥ R T |x|=\varepsilon_sx\ge R_T ∣ x ∣ = ε s x ≥ R T . Assumption 3.2 and ∣ y ∣ ≥ 1 |y|\ge1 ∣ y ∣ ≥ 1 therefore give
s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ 1 f 1 , s ) ≤ C T ( 1 + ( 2 C ) 1 f 1 , s ∣ y ∣ ) ≤ C T ( 1 + ( 2 C ) 1 f 1 , s ) ∣ y ∣ . s\rho\Gamma_1(t,x)\sigma(t,x)\le C_T\left(1+|x|^{\frac{1}{f_{1,s}}}\right)\le C_T\left(1+(2C)^{\frac{1}{f_{1,s}}}|y|\right)\le C_T\left(1+(2C)^{\frac{1}{f_{1,s}}}\right)|y|. s ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ( 1 + ∣ x ∣ f 1 , s 1 ) ≤ C T ( 1 + ( 2 C ) f 1 , s 1 ∣ y ∣ ) ≤ C T ( 1 + ( 2 C ) f 1 , s 1 ) ∣ y ∣.
Since ( y + ) f 1 , + + ( y − ) f 1 , − = ∣ y ∣ f 1 , s (y^+)^{f_{1,+}}+(y^-)^{f_{1,-}}=|y|^{f_{1,s}} ( y + ) f 1 , + + ( y − ) f 1 , − = ∣ y ∣ f 1 , s when s y > 0 sy>0 sy > 0 , a further increase of C T C_T C T yields, for all ( t , y , w ) ∈ [ 0 , T ] × R 2 (t,y,w)\in[0,T]\times\mathbb{R}^2 ( t , y , w ) ∈ [ 0 , T ] × R 2 ,
( y + ) f 1 , + + ( y − ) f 1 , − ≥ C T ( 1 + ∣ w ∣ f 2 ) ⟹ sgn ( y ) ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ∣ y ∣ . (33) (y^+)^{f_{1,+}}+(y^-)^{f_{1,-}}\ge C_T(1+|w|^{f_2})\quad\Longrightarrow\quad\operatorname{sgn}(y)\rho\Gamma_1(t,x)\sigma(t,x)\le C_T|y|. \tag*{(33)} ( y + ) f 1 , + + ( y − ) f 1 , − ≥ C T ( 1 + ∣ w ∣ f 2 ) ⟹ sgn ( y ) ρ Γ 1 ( t , x ) σ ( t , x ) ≤ C T ∣ y ∣. (33)
Step 1: Bound for Z + Z^+ Z + . We fix n ∈ N ∗ n\in\mathbb{N}^* n ∈ N ∗ and t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] . For each trajectory, we set
m t : = 1 + sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ f 1 + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 . m_t:=1+\sup_{u\in[0,t]}|U^n_{1,u}|^{f_1}+\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{f_2}. m t := 1 + u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ f 1 + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 2 .
With C T C_T C T as in (33) , we set
A T : = 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + max { C T 1 f 1 , + , C T 1 f 1 , − } . A_T := 1+\lVert G_1\rVert_{\infty,[0,T]}+\max\left\{C_T^{\frac{1}{f_{1,+}}},C_T^{\frac{1}{f_{1,-}}}\right\}. A T := 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + max { C T f 1 , + 1 , C T f 1 , − 1 } .
On [ 0 , t ∧ τ n ] [0,t\wedge\tau_n] [ 0 , t ∧ τ n ] , we have Z = U 1 n + B Z=U_1^n+B Z = U 1 n + B , with
∀ s ∈ [ 0 , t ∧ τ n ] , B s = ρ ∫ 0 s K 1 ( s − u ) Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) d u . \forall s\in[0,t\wedge\tau_n],\qquad B_s=\rho\int_0^s K_1(s-u)\Gamma_1(u,\widetilde X_u)\sigma(u,\widetilde X_u)\,\mathrm{d}u. ∀ s ∈ [ 0 , t ∧ τ n ] , B s = ρ ∫ 0 s K 1 ( s − u ) Γ 1 ( u , X u ) σ ( u , X u ) d u .
Since m t ≥ 1 m_t\ge1 m t ≥ 1 and f 1 , + ≤ f 1 f_{1,+}\le f_1 f 1 , + ≤ f 1 , we have
∣ U 1 , s n ∣ f 1 , + ≤ 1 + ∣ U 1 , s n ∣ f 1 ≤ m t , \lvert U_{1,s}^n\rvert^{f_{1,+}}\le1+\lvert U_{1,s}^n\rvert^{f_1}\le m_t, ∣ U 1 , s n ∣ f 1 , + ≤ 1 + ∣ U 1 , s n ∣ f 1 ≤ m t ,
so that
∀ s ∈ [ 0 , t ∧ τ n ] , ∣ U 1 , s n ∣ ∨ ∣ Y ~ 2 , s ∣ f 2 f 1 , + ≤ m t 1 f 1 , + . (34) \forall s\in[0,t\wedge\tau_n],\qquad\lvert U_{1,s}^n\rvert\vee\lvert\widetilde Y_{2,s}\rvert^{\frac{f_2}{f_{1,+}}}\le m_t^{\frac{1}{f_{1,+}}}. \tag*{(34)} ∀ s ∈ [ 0 , t ∧ τ n ] , ∣ U 1 , s n ∣ ∨ ∣ Y 2 , s ∣ f 1 , + f 2 ≤ m t f 1 , + 1 . (34)
If B s > A T m t 1 f 1 , + B_s>A_Tm_t^{\frac{1}{f_{1,+}}} B s > A T m t f 1 , + 1 , then, using m t ≥ 1 m_t\ge1 m t ≥ 1 , (34) , and the definition of A T A_T A T , Y ~ 1 , s = G 1 ( s ) + U 1 , s n + B s > C T 1 f 1 , + m t 1 f 1 , + > 0 \widetilde Y_{1,s}=G_1(s)+U_{1,s}^n+B_s>C_T^{\frac{1}{f_{1,+}}}m_t^{\frac{1}{f_{1,+}}}>0 Y 1 , s = G 1 ( s ) + U 1 , s n + B s > C T f 1 , + 1 m t f 1 , + 1 > 0 . Consequently, the definition of m t m_t m t gives
( Y ~ 1 , s ) f 1 , + > C T m t ≥ C T ( 1 + ∣ Y ~ 2 , s ∣ f 2 ) . \left(\widetilde Y_{1,s}\right)^{f_{1,+}}>C_Tm_t\ge C_T\left(1+\lvert\widetilde Y_{2,s}\rvert^{f_2}\right). ( Y 1 , s ) f 1 , + > C T m t ≥ C T ( 1 + ∣ Y 2 , s ∣ f 2 ) .
Thus, if B s > A T m t 1 f 1 , + B_s>A_Tm_t^{\frac{1}{f_{1,+}}} B s > A T m t f 1 , + 1 , the drift bound (33) , with x = X ~ s = β ( s ) + F 1 ( Y ~ 1 , s ) + F 2 ( Y ~ 2 , s ) x=\widetilde X_s=\beta(s)+F_1(\widetilde Y_{1,s})+F_2(\widetilde Y_{2,s}) x = X s = β ( s ) + F 1 ( Y 1 , s ) + F 2 ( Y 2 , s ) , y = Y ~ 1 , s y=\widetilde Y_{1,s} y = Y 1 , s and w = Y ~ 2 , s w=\widetilde Y_{2,s} w = Y 2 , s , gives
ρ Γ 1 ( s , X ~ s ) σ ( s , X ~ s ) m t 1 f 1 , + ≤ C T Y ~ 1 , s m t 1 f 1 , + ≤ C T ( 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + B s m t 1 f 1 , + ) ≤ C T ( 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + A T ) [ 1 + ( B s m t 1 f 1 , + − A T ) + ] , (35) \begin{aligned}
\frac{\rho\Gamma_1(s,\widetilde X_s)\sigma(s,\widetilde X_s)}{m_t^{\frac{1}{f_{1,+}}}}
\le C_T\frac{\widetilde Y_{1,s}}{m_t^{\frac{1}{f_{1,+}}}} \\
&\le C_T\left(1+\lVert G_1\rVert_{\infty,[0,T]}+\frac{B_s}{m_t^{\frac{1}{f_{1,+}}}}\right) \\
&\le C_T\left(1+\lVert G_1\rVert_{\infty,[0,T]}+A_T\right)\left[1+\left(\frac{B_s}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\right], \tag*{(35)}
\end{aligned} m t f 1 , + 1 ρ Γ 1 ( s , X s ) σ ( s , X s ) ≤ C T m t f 1 , + 1 Y 1 , s ≤ C T 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + m t f 1 , + 1 B s ≤ C T ( 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + A T ) 1 + m t f 1 , + 1 B s − A T + , (35)
where we again used m t ≥ 1 m_t\ge1 m t ≥ 1 and (34) . Continuity of Γ 1 σ \Gamma_1\sigma Γ 1 σ and X ~ \widetilde X X gives
sup s ∈ [ 0 , t ∧ τ n ] ∣ Γ 1 ( s , X ~ s ) σ ( s , X ~ s ) ∣ < ∞ . \sup_{s\in[0,t\wedge\tau_n]}\lvert\Gamma_1(s,\widetilde X_s)\sigma(s,\widetilde X_s)\rvert<\infty. s ∈ [ 0 , t ∧ τ n ] sup ∣ Γ 1 ( s , X s ) σ ( s , X s )∣ < ∞.
For 1 < p ≤ 2 1<p\le2 1 < p ≤ 2 , Lemma 4.2 therefore applies pathwise on [ 0 , t ∧ τ n ] [0,t\wedge\tau_n] [ 0 , t ∧ τ n ] with
∀ x ∈ R , Φ ( x ) = [ ( x m t 1 f 1 , + − A T ) + ] p , \forall x\in\mathbb{R},\qquad\Phi(x)=\left[\left(\frac{x}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\right]^p, ∀ x ∈ R , Φ ( x ) = m t f 1 , + 1 x − A T + p ,
∀ s ∈ [ 0 , t ∧ τ n ] , g s = ρ Γ 1 ( s , X ~ s ) σ ( s , X ~ s ) , \forall s\in[0,t\wedge\tau_n],\qquad g_s=\rho\Gamma_1(s,\widetilde X_s)\sigma(s,\widetilde X_s), ∀ s ∈ [ 0 , t ∧ τ n ] , g s = ρ Γ 1 ( s , X s ) σ ( s , X s ) ,
because, for p > 1 p>1 p > 1 , Φ \Phi Φ belongs to C 1 ( R ) C^1(\mathbb{R}) C 1 ( R ) and is convex and nonnegative, with Φ ( 0 ) = 0 \Phi(0)=0 Φ ( 0 ) = 0 . Since K 1 ∈ L 1 ( [ 0 , T ] ) K_1\in L^1([0,T]) K 1 ∈ L 1 ([ 0 , T ]) , B B B is bounded on [ 0 , t ∧ τ n ] [0,t\wedge\tau_n] [ 0 , t ∧ τ n ] . Moreover, we have
Φ ′ ( B s ) = p m t 1 f 1 , + [ ( B s m t 1 f 1 , + − A T ) + ] p − 1 ⟶ p ↓ 1 1 { B s > A T m t 1 f 1 , + } m t 1 f 1 , + . \Phi'(B_s)=\frac{p}{m_t^{\frac{1}{f_{1,+}}}}\left[\left(\frac{B_s}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\right]^{p-1}\longrightarrow_{p\downarrow1}\frac{\mathbf{1}_{\{B_s>A_Tm_t^{\frac{1}{f_{1,+}}}\}}}{m_t^{\frac{1}{f_{1,+}}}}. Φ ′ ( B s ) = m t f 1 , + 1 p m t f 1 , + 1 B s − A T + p − 1 ⟶ p ↓ 1 m t f 1 , + 1 1 { B s > A T m t f 1 , + 1 } .
Since z p p − 1 ≤ 2 ( 1 + z ) zp^{p-1}\leq2(1+z) z p p − 1 ≤ 2 ( 1 + z ) for z ≥ 0 z\geq0 z ≥ 0 and 1 < p ≤ 2 1<p\leq2 1 < p ≤ 2 , boundedness of B , g B,g B , g gives an integrable bound after multiplication by K 1 ( u − s ) K_{1}(u-s) K 1 ( u − s ) . Dominated convergence as p ↓ 1 p\downarrow1 p ↓ 1 , together with the drift bound (35) , then gives for all u ∈ [ 0 , t ∧ τ n ] u\in[0,t\wedge\tau_n] u ∈ [ 0 , t ∧ τ n ] ,
( B u m t 1 f 1 , + − A T ) + ≤ ∫ 0 u K 1 ( u − s ) 1 { B s > A T m t 1 f 1 , + } ρ Γ 1 ( s , X ~ s ) σ ( s , X ~ s ) m t 1 f 1 , + d s ≤ C T ( 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + A T ) ∫ 0 u K 1 ( u − s ) [ 1 + ( B s m t 1 f 1 , + − A T ) + ] d s . (36) \begin{aligned}
\left(\frac{B_u}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+
\leq\int_0^u K_1(u-s)\mathbf{1}_{\left\{B_s>A_Tm_t^{\frac{1}{f_{1,+}}}\right\}}\frac{\rho\Gamma_1(s,\widetilde X_s)\sigma(s,\widetilde X_s)}{m_t^{\frac{1}{f_{1,+}}}}\,ds \\
&\leq C_T(1+\|G_1\|_{\infty,[0,T]}+A_T)\int_0^u K_1(u-s)\left[1+\left(\frac{B_s}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\right]\,ds. \tag*{(36)}
\end{aligned} m t f 1 , + 1 B u − A T + ≤ ∫ 0 u K 1 ( u − s ) 1 { B s > A T m t f 1 , + 1 } m t f 1 , + 1 ρ Γ 1 ( s , X s ) σ ( s , X s ) d s ≤ C T ( 1 + ∥ G 1 ∥ ∞ , [ 0 , T ] + A T ) ∫ 0 u K 1 ( u − s ) 1 + m t f 1 , + 1 B s − A T + d s . (36)
The Volterra–Grönwall lemma [37 Lemma 2.2] gives
sup u ∈ [ 0 , t ∧ τ n ] ( B u m t 1 f 1 , + − A T ) + ≤ C T . \sup_{u\in[0,t\wedge\tau_n]}\left(\frac{B_u}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\leq C_T. u ∈ [ 0 , t ∧ τ n ] sup m t f 1 , + 1 B u − A T + ≤ C T .
Since for all u ∈ [ 0 , t ∧ τ n ] u\in[0,t\wedge\tau_n] u ∈ [ 0 , t ∧ τ n ] ,
Z u + = ( U 1 , u n + B u ) + ≤ ( m t 1 f 1 , + + B u ) + ≤ m t 1 f 1 , + [ 1 + A T + ( B u m t 1 f 1 , + − A T ) + ] , Z_u^+=(U_{1,u}^n+B_u)^+\leq\left(m_t^{\frac{1}{f_{1,+}}}+B_u\right)^+\leq m_t^{\frac{1}{f_{1,+}}}\left[1+A_T+\left(\frac{B_u}{m_t^{\frac{1}{f_{1,+}}}}-A_T\right)^+\right], Z u + = ( U 1 , u n + B u ) + ≤ ( m t f 1 , + 1 + B u ) + ≤ m t f 1 , + 1 1 + A T + m t f 1 , + 1 B u − A T + ,
we obtain
sup u ∈ [ 0 , t ] ( Z u ∧ τ n + ) f 1 , + ≤ C T m t . (37) \sup_{u\in[0,t]}(Z_{u\wedge\tau_n}^+)^{f_{1,+}}\leq C_Tm_t. \tag*{(37)} u ∈ [ 0 , t ] sup ( Z u ∧ τ n + ) f 1 , + ≤ C T m t . (37)
Step 2: Bound for X ~ \widetilde X X . The argument for Z − Z^- Z − is obtained by replacing ( Z , B , g , U 1 n , G 1 , Y ~ 1 ) (Z,B,g,U_1^n,G_1,\widetilde Y_1) ( Z , B , g , U 1 n , G 1 , Y 1 ) by ( − Z , − B , − g , − U 1 n , − G 1 , − Y ~ 1 ) (-Z,-B,-g,-U_1^n,-G_1,-\widetilde Y_1) ( − Z , − B , − g , − U 1 n , − G 1 , − Y 1 ) and f 1 , + f_{1,+} f 1 , + by f 1 , − f_{1,-} f 1 , − in the proof of (37) . The required drift bound follows from (33) with y = Y ~ 1 , s < 0 y=\widetilde Y_{1,s}<0 y = Y 1 , s < 0 . Thus, with the same m t m_t m t , we have
sup u ∈ [ 0 , t ] ( Z u ∧ τ n − ) f 1 , − ≤ C T m t . \sup_{u\in[0,t]}(Z_{u\wedge\tau_n}^-)^{f_{1,-}}\leq C_Tm_t. u ∈ [ 0 , t ] sup ( Z u ∧ τ n − ) f 1 , − ≤ C T m t .
Moreover, Y ~ 1 = G 1 + Z \widetilde Y_1=G_1+Z Y 1 = G 1 + Z gives
( Y ~ 1 , u ) + ≤ ∥ G 1 ∥ ∞ , [ 0 , T ] + Z u + . (\widetilde Y_{1,u})^+\leq\|G_1\|_{\infty,[0,T]}+Z_u^+. ( Y 1 , u ) + ≤ ∥ G 1 ∥ ∞ , [ 0 , T ] + Z u + .
The tail bounds on F 1 F_1 F 1 and its continuity therefore yield
∣ F 1 ( Y ~ 1 , u ) ∣ ≤ C T ( 1 + ( Z u + ) f 1 , + + ( Z u − ) f 1 , − ) . |F_1(\widetilde Y_{1,u})|\leq C_T\left(1+(Z_u^+)^{f_{1,+}}+(Z_u^-)^{f_{1,-}}\right). ∣ F 1 ( Y 1 , u ) ∣ ≤ C T ( 1 + ( Z u + ) f 1 , + + ( Z u − ) f 1 , − ) .
Since β , G 1 \beta,G_1 β , G 1 are bounded on [ 0 , T ] [0,T] [ 0 , T ] , the polynomial growth bounds on F 1 , F 2 F_1,F_2 F 1 , F 2 in Assumption 1.2 give
sup u ∈ [ 0 , t ] ∣ X ~ u ∧ τ n ∣ ≤ C T ( 1 + sup u ∈ [ 0 , t ] ( ( Z u ∧ τ n + ) f 1 , + + ( Z u ∧ τ n − ) f 1 , − ) + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 ) ≤ C T m t . (38) \begin{aligned}
\sup_{u\in[0,t]}|\widetilde X_{u\wedge\tau_n}|
\leq C_T\left(1+\sup_{u\in[0,t]}\left((Z_{u\wedge\tau_n}^+)^{f_{1,+}}+(Z_{u\wedge\tau_n}^-)^{f_{1,-}}\right)+\sup_{u\in[0,t]}|\widetilde Y_{2,u\wedge\tau_n}|^{f_2}\right) \\
&\leq C_Tm_t. \tag*{(38)}
\end{aligned} u ∈ [ 0 , t ] sup ∣ X u ∧ τ n ∣ ≤ C T ( 1 + u ∈ [ 0 , t ] sup ( ( Z u ∧ τ n + ) f 1 , + + ( Z u ∧ τ n − ) f 1 , − ) + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 2 ) ≤ C T m t . (38)
Hence (27) of Lemma 4.1 holds, which proves martingality on [ 0 , T ] [0,T] [ 0 , T ] . Since T > 0 T>0 T > 0 is arbitrary, ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 is a martingale.
Proof of Theorem 1(B.i)
Proof of Theorem 1(B.i). We fix s ∈ { − 1 , 1 } s \in\{-1,1\} s ∈ { − 1 , 1 } as in Assumption 3.1. We use U n U^n U n as defined in Lemma 4.1. The constant L T L_T L T is the upper bound for ε s X \varepsilon_s X ε s X in Assumption 3.1. For every n ∈ N ∗ n \in\mathbb{N}^* n ∈ N ∗ , the density of Q n \mathbb{Q}_n Q n , defined in the proof of Lemma 3.1 with θ = 0 \theta=0 θ = 0 , is strictly positive. Thus Q n \mathbb{Q}_n Q n and P 0 \mathbb{P}_0 P 0 are equivalent on F T \mathcal{F}_T F T , and the upper bound on ε s X \varepsilon_s X ε s X in Assumption 3.1 holds Q n \mathbb{Q}_n Q n -almost surely. The equality of the stopped laws in (26) , with θ = 0 \theta=0 θ = 0 , gives
P 0 ( sup t ∈ [ 0 , T ] ε s X ~ t ∧ τ n ≤ L T ) = 1. (39) \mathbb{P}_0\left(\sup_{t\in[0,T]}\varepsilon_s\widetilde{X}_{t\wedge\tau_n}\leq L_T\right)=1. \tag*{(39)} P 0 ( t ∈ [ 0 , T ] sup ε s X t ∧ τ n ≤ L T ) = 1. (39)
By the bound on − s ρ Γ 1 σ -s\rho\Gamma_1\sigma − s ρ Γ 1 σ in Assumption 3.1, the upper bound (39) , and K 1 ≥ 0 K_1\geq0 K 1 ≥ 0 , we have
∀ t ∈ [ 0 , T ∧ τ n ] , ( s Y ~ 1 , t ) − ≤ ∣ G 1 ( t ) ∣ + ∣ U 1 , t n ∣ + ∫ 0 t K 1 ( t − u ) ( s ρ Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) ) − d u ≤ ∣ U 1 , t n ∣ + C T ( 1 + ∫ 0 t K 1 ( t − u ) ∣ X ~ u ∣ 1 f 1 , − s d u ) . (40) \begin{aligned}
\forall t\in[0,T\wedge\tau_n],\qquad(s\widetilde{Y}_{1,t})^- \leq|G_1(t)|+|U^n_{1,t}|+\int_0^t K_1(t-u)(s\rho\Gamma_1(u,\widetilde{X}_u)\sigma(u,\widetilde{X}_u))^-\,\mathrm{d}u \\
&\leq|U^n_{1,t}|+C_T\left(1+\int_0^t K_1(t-u)|\widetilde{X}_u|^{\frac{1}{f_{1,-s}}}\,\mathrm{d}u\right). \tag*{(40)}
\end{aligned} ∀ t ∈ [ 0 , T ∧ τ n ] , ( s Y 1 , t ) − ≤ ∣ G 1 ( t ) ∣ + ∣ U 1 , t n ∣ + ∫ 0 t K 1 ( t − u ) ( s ρ Γ 1 ( u , X u ) σ ( u , X u ) ) − d u ≤ ∣ U 1 , t n ∣ + C T ( 1 + ∫ 0 t K 1 ( t − u ) ∣ X u ∣ f 1 , − s 1 d u ) . (40)
since G 1 G_1 G 1 is bounded on [ 0 , T ] [0,T] [ 0 , T ] . Since ε s F 1 ( y ) = ε s F 1 ( s ( s y ) ) \varepsilon_sF_1(y)=\varepsilon_sF_1(s(sy)) ε s F 1 ( y ) = ε s F 1 ( s ( sy )) , the bounds on F 1 F_1 F 1 in Assumption 1.2 give
∀ y ∈ R , ( ε s F 1 ( y ) ) − ≤ C ( 1 + ( ( s y ) − ) f 1 , − s ) . \forall y\in\mathbb{R},\qquad(\varepsilon_sF_1(y))^- \leq C\left(1+((sy)^-)^{f_{1,-s}}\right). ∀ y ∈ R , ( ε s F 1 ( y ) ) − ≤ C ( 1 + (( sy ) − ) f 1 , − s ) .
Indeed, this negative part vanishes when s y ≥ y ∗ sy\geq y_* sy ≥ y ∗ , by Assumption 1.2, and is bounded on 0 ≤ s y ≤ y ∗ 0\leq sy\leq y_* 0 ≤ sy ≤ y ∗ by continuity of F 1 F_1 F 1 . When s y < 0 sy<0 sy < 0 , the growth bound in the direction − s -s − s gives
( ε s F 1 ( y ) ) − ≤ ∣ F 1 ( y ) ∣ = ∣ F 1 ( − s ( − s y ) ) ∣ ≤ C ( 1 + ( − s y ) f 1 , − s ) . (\varepsilon_sF_1(y))^- \leq|F_1(y)|=|F_1(-s(-sy))|\leq C\left(1+(-sy)^{f_{1,-s}}\right). ( ε s F 1 ( y ) ) − ≤ ∣ F 1 ( y ) ∣ = ∣ F 1 ( − s ( − sy )) ∣ ≤ C ( 1 + ( − sy ) f 1 , − s ) .
The upper bound (39) therefore gives
∀ t ∈ [ 0 , T ∧ τ n ] , ∣ X ~ t ∣ = ε s X ~ t + 2 ( ε s X ~ t ) − ≤ ∣ L T ∣ + 2 ∣ β ( t ) ∣ + 2 ( ε s F 1 ( Y ~ 1 , t ) ) − + 2 ∣ F 2 ( Y ~ 2 , t ) ∣ ≤ C T ( 1 + ( ( s Y ~ 1 , t ) − ) f 1 , − s + ∣ Y ~ 2 , t ∣ f 2 ) . \begin{aligned}
\forall t\in[0,T\wedge\tau_n],\qquad|\widetilde{X}_t|=\varepsilon_s\widetilde{X}_t+2(\varepsilon_s\widetilde{X}_t)^- \\
&\leq|L_T|+2|\beta(t)|+2(\varepsilon_sF_1(\widetilde{Y}_{1,t}))^-+2|F_2(\widetilde{Y}_{2,t})| \\
&\leq C_T\left(1+((s\widetilde{Y}_{1,t})^-)^{f_{1,-s}}+|\widetilde{Y}_{2,t}|^{f_2}\right).
\end{aligned} ∀ t ∈ [ 0 , T ∧ τ n ] , ∣ X t ∣ = ε s X t + 2 ( ε s X t ) − ≤ ∣ L T ∣ + 2∣ β ( t ) ∣ + 2 ( ε s F 1 ( Y 1 , t ) ) − + 2∣ F 2 ( Y 2 , t ) ∣ ≤ C T ( 1 + (( s Y 1 , t ) − ) f 1 , − s + ∣ Y 2 , t ∣ f 2 ) .
Raising both sides to the power 1 f 1 , − s \frac{1}{f_{1,-s}} f 1 , − s 1 and using the bound (40) on ( s Y ~ 1 , t ) − (s\widetilde{Y}_{1,t})^- ( s Y 1 , t ) − , we obtain
∀ t ∈ [ 0 , T ∧ τ n ] , ∣ X ~ t ∣ 1 f 1 , − s ≤ C T ( 1 + ∣ U 1 , t n ∣ + ∣ Y ~ 2 , t ∣ f 2 f 1 , − s + ∫ 0 t K 1 ( t − u ) ∣ X ~ u ∣ 1 f 1 , − s d u ) . \forall t\in[0,T\wedge\tau_n],\qquad|\widetilde{X}_t|^{\frac{1}{f_{1,-s}}}\leq C_T\left(1+|U^n_{1,t}|+|\widetilde{Y}_{2,t}|^{\frac{f_2}{f_{1,-s}}}+\int_0^t K_1(t-u)|\widetilde{X}_u|^{\frac{1}{f_{1,-s}}}\,\mathrm{d}u\right). ∀ t ∈ [ 0 , T ∧ τ n ] , ∣ X t ∣ f 1 , − s 1 ≤ C T ( 1 + ∣ U 1 , t n ∣ + ∣ Y 2 , t ∣ f 1 , − s f 2 + ∫ 0 t K 1 ( t − u ) ∣ X u ∣ f 1 , − s 1 d u ) .
For each trajectory and t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , the Volterra–Grönwall lemma [37 Lemma 2.2] , applied on [ 0 , t ∧ τ n ] [0,t\wedge\tau_n] [ 0 , t ∧ τ n ] , gives
sup u ∈ [ 0 , t ] ∣ X ~ u ∧ τ n ∣ 1 f 1 , − s ≤ C T ( 1 + sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 f 1 , − s ) . \sup_{u\in[0,t]}|\widetilde{X}_{u\wedge\tau_n}|^{\frac{1}{f_{1,-s}}}\leq C_T\left(1+\sup_{u\in[0,t]}|U^n_{1,u}|+\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{\frac{f_2}{f_{1,-s}}}\right). u ∈ [ 0 , t ] sup ∣ X u ∧ τ n ∣ f 1 , − s 1 ≤ C T ( 1 + u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 1 , − s f 2 ) .
Raising this bound to the power f 1 , − s f_{1,-s} f 1 , − s and using f 1 , − s ≤ f 1 f_{1,-s}\leq f_1 f 1 , − s ≤ f 1 gives
∀ t ∈ [ 0 , T ] , sup u ∈ [ 0 , t ] ∣ X ~ u ∧ τ n ∣ ≤ C T ( 1 + sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ f 1 , − s + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 ) ≤ C T ( 1 + sup u ∈ [ 0 , t ] ∣ U 1 , u n ∣ f 1 + sup u ∈ [ 0 , t ] ∣ Y ~ 2 , u ∧ τ n ∣ f 2 ) . (41) \begin{aligned}
\forall t\in[0,T],\qquad\sup_{u\in[0,t]}|\widetilde{X}_{u\wedge\tau_n}|\leq C_T\left(1+\sup_{u\in[0,t]}|U^n_{1,u}|^{f_{1,-s}}+\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{f_2}\right) \\
&\leq C_T\left(1+\sup_{u\in[0,t]}|U^n_{1,u}|^{f_1}+\sup_{u\in[0,t]}|\widetilde{Y}_{2,u\wedge\tau_n}|^{f_2}\right). \tag*{(41)}
\end{aligned} ∀ t ∈ [ 0 , T ] , u ∈ [ 0 , t ] sup ∣ X u ∧ τ n ∣ ≤ C T ( 1 + u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ f 1 , − s + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 2 ) ≤ C T ( 1 + u ∈ [ 0 , t ] sup ∣ U 1 , u n ∣ f 1 + u ∈ [ 0 , t ] sup ∣ Y 2 , u ∧ τ n ∣ f 2 ) . (41)
Hence (27) of Lemma 4.1 holds, which proves martingality on [ 0 , T ] [0,T] [ 0 , T ] . Since T > 0 T>0 T > 0 is arbitrary, ( S t ) t ≥ 0 (S_t)_{t\geq0} ( S t ) t ≥ 0 is a martingale. □
Strict local martingality: Proof of Theorem 2
We fix T > 0 T>0 T > 0 , a bounded stopping time θ \theta θ , and a deterministic T ‾ < ∞ \overline{T}<\infty T < ∞ such that θ + T ≤ T ‾ \theta+T\leq\overline{T} θ + T ≤ T a.s. We write ( X ~ , Y ~ ) = ( X ~ θ , Y ~ θ ) (\widetilde{X},\widetilde{Y})=(\widetilde{X}^{\theta},\widetilde{Y}^{\theta}) ( X , Y ) = ( X θ , Y θ ) for the solution of ( P D V θ ) (\mathrm{PDV}_{\theta}) ( PDV θ ) . Throughout this section, s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } is one possible sign in Assumption 5.1, and ε s \varepsilon_s ε s is fixed by Assumption 1.2.
Outline of the proof and preliminary lemmas
Notation 3. We set
∀ t ∈ [ 0 , + ∞ ) , Z ‾ t : = s ( Y ~ 1 , t − G 1 ( t ) ) . \forall t\in[0,+\infty),\qquad\overline{Z}_t:=s(\widetilde{Y}_{1,t}-G_1(t)). ∀ t ∈ [ 0 , + ∞ ) , Z t := s ( Y 1 , t − G 1 ( t )) .
By Assumption 2.2, X ~ \widetilde{X} X and Y ~ \widetilde{Y} Y are continuous and adapted before explosion. Since G 1 G_1 G 1 is continuous, the same holds for Z ‾ \overline{Z} Z . The first hitting and exit times used below are therefore stopping times. The proof is summarized in Figure 2.
Figure 2. Structure of the explosion proof.
We use the resolvent of K 1 K_1 K 1 to obtain a semimartingale representation of Z ‾ \overline{Z} Z . The bound in Lemma 5.1(ii) below controls the memory drift term in (5.2). Following [10 ] , we estimate the memory drift term after stopping the solution. Under the kernel regularity in Assumption 4.2, R 1 \mathcal{R}_1 R 1 denotes the resolvent of the first kind constructed in Appendix § 9 , and
r m e m : = K 1 ′ ∗ R 1 ∈ B V l o c ( R + ) . (42) r_{\mathrm{mem}}:=K'_1*\mathcal{R}_1\in BV_{\mathrm{loc}}(\mathbb{R}_+). \tag*{(42)} r mem := K 1 ′ ∗ R 1 ∈ B V loc ( R + ) . (42)
We define the signed measure d r m e m dr_{\mathrm{mem}} d r mem by d r m e m ( ( a , b ] ) : = r m e m ( b ) − r m e m ( a ) dr_{\mathrm{mem}}((a,b]):=r_{\mathrm{mem}}(b)-r_{\mathrm{mem}}(a) d r mem (( a , b ]) := r mem ( b ) − r mem ( a ) for 0 ≤ a < b 0\leq a<b 0 ≤ a < b , with no atom at zero; ∣ d r m e m ∣ |dr_{\mathrm{mem}}| ∣ d r mem ∣ denotes its total variation.
Lemma 5.1. Suppose that Assumption 4.2 holds. For R ∈ N ∗ R\in\mathbb{N}^* R ∈ N ∗ , we set
τ R X ~ : = inf { t ≥ 0 : ∣ X ~ t ∣ ≥ R } . \tau^{\widetilde{X}}_R:=\inf\{t\geq0:|\widetilde{X}_t|\geq R\}. τ R X := inf { t ≥ 0 : ∣ X t ∣ ≥ R } .
We use R 1 \mathcal{R}_1 R 1 from (B.2) and r m e m r_{\mathrm{mem}} r mem from (5.1). Then the following statements hold:
(i) The stopped process ( Z ‾ t ∧ τ R X ~ ) t ≥ 0 (\overline{Z}_{t\wedge\tau^{\widetilde{X}}_R})_{t\geq0} ( Z t ∧ τ R X ) t ≥ 0 is a semimartingale and satisfies
∀ t ∈ [ 0 , τ R X ~ ] , Z ‾ t = K 1 ( 0 ) ( s ∫ 0 t Γ 1 ( u , X ~ u ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) d u ) + ∫ 0 t ( r m e m ( 0 ) Z ‾ u + ( d r m e m ∗ Z ‾ ) ( u ) ) d u . (43) \forall t\in[0,\tau^{\widetilde{X}}_R],\qquad
\overline{Z}_t=K_1(0)\left(s\int_0^t\Gamma_1(u,\widetilde{X}_u)\,dW_u+s\rho\int_0^t\mathbf{1}_{\{u\geq\theta\}}\Gamma_1(u,\widetilde{X}_u)\sigma(u,\widetilde{X}_u)\,du\right)
+\int_0^t\left(r_{\mathrm{mem}}(0)\overline{Z}_u+(dr_{\mathrm{mem}}*\overline{Z})(u)\right)\,du. \tag*{(43)} ∀ t ∈ [ 0 , τ R X ] , Z t = K 1 ( 0 ) ( s ∫ 0 t Γ 1 ( u , X u ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } Γ 1 ( u , X u ) σ ( u , X u ) d u ) + ∫ 0 t ( r mem ( 0 ) Z u + ( d r mem ∗ Z ) ( u ) ) d u . (43)
@page 24/38 starts_mid=0 ends_mid=1
Strict local martingality: Proof of Theorem 2
We fix T > 0 T>0 T > 0 , a bounded stopping time θ \theta θ , and a deterministic T ‾ < ∞ \overline{T}<\infty T < ∞ such that θ + T ≤ T ‾ \theta+T\leq\overline{T} θ + T ≤ T a.s. We write ( X ~ , Y ~ ) = ( X ~ θ , Y ~ θ ) (\widetilde{X},\widetilde{Y})=(\widetilde{X}^{\theta},\widetilde{Y}^{\theta}) ( X , Y ) = ( X θ , Y θ ) for the solution of ( P D V θ ) (\mathrm{PDV}_{\theta}) ( PDV θ ) . Throughout this section, s ∈ { − 1 , 1 } s\in\{-1,1\} s ∈ { − 1 , 1 } is one possible sign in Assumption 5.1, and ε s \varepsilon_s ε s is fixed by Assumption 1.2.
Outline of the proof and preliminary lemmas
Notation 3. We set
∀ t ∈ [ 0 , + ∞ ) , Z ‾ t : = s ( Y ~ 1 , t − G 1 ( t ) ) . \forall t\in[0,+\infty),\qquad\overline{Z}_t:=s(\widetilde{Y}_{1,t}-G_1(t)). ∀ t ∈ [ 0 , + ∞ ) , Z t := s ( Y 1 , t − G 1 ( t )) .
By Assumption 2.2, X ~ \widetilde{X} X and Y ~ \widetilde{Y} Y are continuous and adapted before explosion. Since G 1 G_1 G 1 is continuous, the same holds for Z ‾ \overline{Z} Z . The first hitting and exit times used below are therefore stopping times. The proof is summarized in Figure 2.
Figure 2 (PDF p. 24). Structure of the explosion proof.
We use the resolvent of K 1 K_1 K 1 to obtain a semimartingale representation of Z ‾ \overline{Z} Z . The bound in Lemma 5.1(ii) below controls the memory drift term in (5.2). Following [10 ] , we estimate the memory drift term after stopping the solution. Under the kernel regularity in Assumption 4.2, R 1 \mathcal{R}_1 R 1 denotes the resolvent of the first kind constructed in Appendix § 9 , and
r m e m : = K 1 ′ ∗ R 1 ∈ B V l o c ( R + ) . (44) r_{\mathrm{mem}}:=K'_1*\mathcal{R}_1\in BV_{\mathrm{loc}}(\mathbb{R}_+). \tag*{(44)} r mem := K 1 ′ ∗ R 1 ∈ B V loc ( R + ) . (44)
We define the signed measure d r m e m dr_{\mathrm{mem}} d r mem by d r m e m ( ( a , b ] ) : = r m e m ( b ) − r m e m ( a ) dr_{\mathrm{mem}}((a,b]):=r_{\mathrm{mem}}(b)-r_{\mathrm{mem}}(a) d r mem (( a , b ]) := r mem ( b ) − r mem ( a ) for 0 ≤ a < b 0\leq a<b 0 ≤ a < b , with no atom at zero; ∣ d r m e m ∣ |dr_{\mathrm{mem}}| ∣ d r mem ∣ denotes its total variation.
Lemma 5.1. Suppose that Assumption 4.2 holds. For R ∈ N ∗ R\in\mathbb{N}^* R ∈ N ∗ , we set
τ R X ~ : = inf { t ≥ 0 : ∣ X ~ t ∣ ≥ R } . \tau^{\widetilde{X}}_R:=\inf\{t\geq0:|\widetilde{X}_t|\geq R\}. τ R X := inf { t ≥ 0 : ∣ X t ∣ ≥ R } .
We use R 1 \mathcal{R}_1 R 1 from (B.2) and r m e m r_{\mathrm{mem}} r mem from (5.1). Then the following statements hold:
(i) The stopped process ( Z ‾ t ∧ τ R X ~ ) t ≥ 0 (\overline{Z}_{t\wedge\tau^{\widetilde{X}}_R})_{t\geq0} ( Z t ∧ τ R X ) t ≥ 0 is a semimartingale and satisfies
∀ t ∈ [ 0 , τ R X ~ ] , Z ‾ t = K 1 ( 0 ) ( s ∫ 0 t Γ 1 ( u , X ~ u ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) d u ) + ∫ 0 t ( r m e m ( 0 ) Z ‾ u + ( d r m e m ∗ Z ‾ ) ( u ) ) d u . (45) \forall t\in[0,\tau^{\widetilde{X}}_R],\qquad
\overline{Z}_t=K_1(0)\left(s\int_0^t\Gamma_1(u,\widetilde{X}_u)\,dW_u+s\rho\int_0^t\mathbf{1}_{\{u\geq\theta\}}\Gamma_1(u,\widetilde{X}_u)\sigma(u,\widetilde{X}_u)\,du\right)
+\int_0^t\left(r_{\mathrm{mem}}(0)\overline{Z}_u+(dr_{\mathrm{mem}}*\overline{Z})(u)\right)\,du. \tag*{(45)} ∀ t ∈ [ 0 , τ R X ] , Z t = K 1 ( 0 ) ( s ∫ 0 t Γ 1 ( u , X u ) d W u + s ρ ∫ 0 t 1 { u ≥ θ } Γ 1 ( u , X u ) σ ( u , X u ) d u ) + ∫ 0 t ( r mem ( 0 ) Z u + ( d r mem ∗ Z ) ( u ) ) d u . (45)
(ii)
∀ t ∈ [ 0 , T ‾ ∧ τ ~ R ] , ∣ r m e m ( 0 ) Z ‾ t + ( d r m e m ∗ Z ‾ ) ( t ) ∣ ≤ ( ∣ r m e m ( 0 ) ∣ + ∣ d r m e m ∣ ( [ 0 , T ‾ ] ) ) sup u ∈ [ 0 , t ] ∣ Z ‾ u ∣ . (46) \forall t \in[0,\overline{T}\wedge\widetilde{\tau}_{R}],\qquad\left|r_{\mathrm{mem}}(0)\overline{Z}_{t}+(dr_{\mathrm{mem}}*\overline{Z})(t)\right|\leq\left(\left|r_{\mathrm{mem}}(0)\right|+\left|dr_{\mathrm{mem}}\right|([0,\overline{T}])\right)\sup_{u\in[0,t]}\left|\overline{Z}_{u}\right|. \tag*{(46)} ∀ t ∈ [ 0 , T ∧ τ R ] , r mem ( 0 ) Z t + ( d r mem ∗ Z ) ( t ) ≤ ( ∣ r mem ( 0 ) ∣ + ∣ d r mem ∣ ([ 0 , T ]) ) u ∈ [ 0 , t ] sup Z u . (46)
The proof is given in Appendix § 9 .
Lemma 5.2. Let Z Z Z be a continuous adapted process, τ \tau τ a stopping time, and a , b a,b a , b predictable processes satisfying
∀ s ∈ [ 0 , + ∞ ) , Z s ∧ τ − Z 0 = ∫ 0 s ∧ τ b u d u + ∫ 0 s ∧ τ a u d W u . \forall s\in[0,+\infty),\qquad Z_{s\wedge\tau}-Z_{0}=\int_{0}^{s\wedge\tau}b_{u}\,du+\int_{0}^{s\wedge\tau}a_{u}\,dW_{u}. ∀ s ∈ [ 0 , + ∞ ) , Z s ∧ τ − Z 0 = ∫ 0 s ∧ τ b u d u + ∫ 0 s ∧ τ a u d W u . We assume that there are constants h > 0 h>0 h > 0 and 0 < c ≤ C < ∞ 0<c\leq C<\infty 0 < c ≤ C < ∞ such that τ ≤ h \tau\leq h τ ≤ h almost surely and
P 0 ( ∀ s ∈ [ 0 , τ ) , ∣ b s ∣ ≤ C , c ≤ ∣ a s ∣ ≤ C ) = 1. \mathbb{P}_{0}\left(\forall s\in[0,\tau),\quad|b_{s}|\leq C,\quad c\leq|a_{s}|\leq C\right)=1. P 0 ( ∀ s ∈ [ 0 , τ ) , ∣ b s ∣ ≤ C , c ≤ ∣ a s ∣ ≤ C ) = 1. Then, for every δ > 0 \delta>0 δ > 0 and every p ∈ C 1 ( [ 0 , h ] ) p\in C^{1}([0,h]) p ∈ C 1 ([ 0 , h ]) with p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 , we have
P 0 ( sup s ∈ [ 0 , τ ] ∣ Z s − Z 0 − p ( s ) ∣ < δ ) > 0. (47) \mathbb{P}_{0}\left(\sup_{s\in[0,\tau]}|Z_{s}-Z_{0}-p(s)|<\delta\right)>0. \tag*{(47)} P 0 ( s ∈ [ 0 , τ ] sup ∣ Z s − Z 0 − p ( s ) ∣ < δ ) > 0. (47) Proof. We define the progressively measurable process u u u by u s : = p ′ ( s ) − b s a s u_{s}:=\frac{p'(s)-b_{s}}{a_{s}} u s := a s p ′ ( s ) − b s for s < τ s<\tau s < τ and u s : = 0 u_{s}:=0 u s := 0 afterwards. It is bounded. Girsanov’s theorem therefore gives a probability measure Q \mathbb{Q} Q equivalent to P 0 \mathbb{P}_{0} P 0 , under which Z s ∧ τ − Z 0 − p ( s ∧ τ ) Z_{s\wedge\tau}-Z_{0}-p(s\wedge\tau) Z s ∧ τ − Z 0 − p ( s ∧ τ ) is a continuous martingale with quadratic variation bounded by C 2 h C^{2}h C 2 h . By the Dambis–Dubins–Schwarz theorem, there is a Brownian motion B B B under Q \mathbb{Q} Q , with an independent extension if necessary, such that
Q ( sup s ∈ [ 0 , τ ] ∣ Z s − Z 0 − p ( s ) ∣ < δ ) ≥ Q ( sup v ∈ [ 0 , C 2 h ] ∣ B v ∣ < δ ) > 0. \mathbb{Q}\left(\sup_{s\in[0,\tau]}|Z_{s}-Z_{0}-p(s)|<\delta\right)\geq\mathbb{Q}\left(\sup_{v\in[0,C^{2}h]}|B_{v}|<\delta\right)>0. Q ( s ∈ [ 0 , τ ] sup ∣ Z s − Z 0 − p ( s ) ∣ < δ ) ≥ Q ( v ∈ [ 0 , C 2 h ] sup ∣ B v ∣ < δ ) > 0. Equivalence of the measures gives (47) of Lemma 5.2 .
The same conclusion holds conditionally after a bounded stopping time θ \theta θ , with strictly positive conditional probability and F θ \mathcal{F}_{\theta} F θ -measurable parameters. The proof uses the shifted Brownian increments and localization of the bounds on h h h , C C C , c − 1 c^{-1} c − 1 , and ∥ p ′ ∥ ∞ \|p'\|_{\infty} ∥ p ′ ∥ ∞ in the change of measure.
Proof of explosion
We work conditionally on F θ \mathcal{F}_{\theta} F θ . Under Assumption 1, for u , v ∈ [ 0 , T ‾ ] u,v\in[0,\overline{T}] u , v ∈ [ 0 , T ] with u ≤ v u\leq v u ≤ v , before explosion, the polynomial growth bound on Γ 2 \Gamma_{2} Γ 2 in Assumption 1.2 and the increment bound on K 2 K_{2} K 2 in (17) of Assumption 1 give
∣ Y ~ 2 , v − Y ~ 2 , u ∣ ≤ ∣ G 2 ( v ) − G 2 ( u ) ∣ + ∫ 0 u ∣ K 2 ( v − r ) − K 2 ( u − r ) ∣ ∣ Γ 2 ( r , X ~ r ) ∣ d r + ∫ u v K 2 ( v − r ) ∣ Γ 2 ( r , X ~ r ) ∣ d r ≤ ∣ G 2 ( v ) − G 2 ( u ) ∣ + C T ‾ ( 1 + sup r ∈ [ 0 , v ] ∣ X ~ r ∣ γ 2 ) ∣ v − u ∣ δ . (48) \begin{aligned}
|\widetilde{Y}_{2,v}-\widetilde{Y}_{2,u}| &\leq|G_{2}(v)-G_{2}(u)| \\
&\quad+\int_{0}^{u}|K_{2}(v-r)-K_{2}(u-r)||\Gamma_{2}(r,\widetilde{X}_{r})|\,dr \\
&\quad+\int_{u}^{v}K_{2}(v-r)|\Gamma_{2}(r,\widetilde{X}_{r})|\,dr \\
&\leq|G_{2}(v)-G_{2}(u)|+C_{\overline{T}}\left(1+\sup_{r\in[0,v]}|\widetilde{X}_{r}|^{\gamma_{2}}\right)|v-u|^{\delta}. \tag*{(48)}
\end{aligned} ∣ Y 2 , v − Y 2 , u ∣ ≤ ∣ G 2 ( v ) − G 2 ( u ) ∣ + ∫ 0 u ∣ K 2 ( v − r ) − K 2 ( u − r ) ∣∣ Γ 2 ( r , X r ) ∣ d r + ∫ u v K 2 ( v − r ) ∣ Γ 2 ( r , X r ) ∣ d r ≤ ∣ G 2 ( v ) − G 2 ( u ) ∣ + C T ( 1 + r ∈ [ 0 , v ] sup ∣ X r ∣ γ 2 ) ∣ v − u ∣ δ . (48)
For θ ≤ u ≤ t ≤ θ + T \theta\leq u\leq t\leq\theta+T θ ≤ u ≤ t ≤ θ + T , the increment bound (48) gives
∣ Y ~ 2 , u ∣ ≤ ∣ Y ~ 2 , θ ∣ + 2 ∥ G 2 ∥ ∞ , [ 0 , T ‾ ] + C ( t − θ ) δ 2 ( 1 + sup r ∈ [ 0 , t ] ∣ X ~ r ∣ γ 2 ) . |\widetilde{Y}_{2,u}|\leq|\widetilde{Y}_{2,\theta}|+2\|G_{2}\|_{\infty,[0,\overline{T}]}+C(t-\theta)^{\delta_{2}}\left(1+\sup_{r\in[0,t]}|\widetilde{X}_{r}|^{\gamma_{2}}\right). ∣ Y 2 , u ∣ ≤ ∣ Y 2 , θ ∣ + 2∥ G 2 ∥ ∞ , [ 0 , T ] + C ( t − θ ) δ 2 ( 1 + r ∈ [ 0 , t ] sup ∣ X r ∣ γ 2 ) .
Since γ 2 f 2 ≤ 1 \gamma_2 f_2 \le1 γ 2 f 2 ≤ 1 , the polynomial growth bound on F 2 F_2 F 2 in Assumption 1.2 gives
sup u ∈ [ θ , t ] ∣ F 2 ( Y ~ 2 , u ) ∣ ≤ B + C ( t − θ ) δ 2 f 2 ( 1 + sup r ∈ [ 0 , t ] ∣ X ~ r ∣ γ 2 f 2 ) ≤ B + 2 C T δ 2 f 2 + C ( t − θ ) δ 2 f 2 sup r ∈ [ 0 , t ] ∣ X ~ r ∣ . \sup_{u\in[\theta,t]} |F_2(\widetilde{Y}_{2,u})| \le B+C(t-\theta)^{\delta_2 f_2}\left(1+\sup_{r\in[0,t]}|\widetilde{X}_r|^{\gamma_2 f_2}\right)
\le B+2CT^{\delta_2 f_2}+C(t-\theta)^{\delta_2 f_2}\sup_{r\in[0,t]}|\widetilde{X}_r|. u ∈ [ θ , t ] sup ∣ F 2 ( Y 2 , u ) ∣ ≤ B + C ( t − θ ) δ 2 f 2 ( 1 + r ∈ [ 0 , t ] sup ∣ X r ∣ γ 2 f 2 ) ≤ B + 2 C T δ 2 f 2 + C ( t − θ ) δ 2 f 2 r ∈ [ 0 , t ] sup ∣ X r ∣.
Here B < ∞ B<\infty B < ∞ depends only on the history up to θ \theta θ and the fixed data, while C C C is independent of that history and of t t t . We may increase B B B below. After increasing B B B to absorb 2 C T δ 2 f 2 2CT^{\delta_2 f_2} 2 C T δ 2 f 2 , we obtain
sup u ∈ [ θ , t ] ∣ F 2 ( Y ~ 2 , u ) ∣ ≤ B + C ( t − θ ) δ 2 f 2 sup r ∈ [ 0 , t ] ∣ X ~ r ∣ . (49) \sup_{u\in[\theta,t]} |F_2(\widetilde{Y}_{2,u})| \le B+C(t-\theta)^{\delta_2 f_2}\sup_{r\in[0,t]}|\widetilde{X}_r|. \tag*{(49)} u ∈ [ θ , t ] sup ∣ F 2 ( Y 2 , u ) ∣ ≤ B + C ( t − θ ) δ 2 f 2 r ∈ [ 0 , t ] sup ∣ X r ∣. (49)
Notation 4. We set ϑ n : = inf { t ≥ θ : Z ‾ t ≥ n } \vartheta_n:=\inf\{t\ge\theta:\overline{Z}_t\ge n\} ϑ n := inf { t ≥ θ : Z t ≥ n } . We also define the event at level n n n
D n : = { ϑ n < + ∞ } ∩ { inf u ∈ [ θ , ϑ n ] Y ~ 1 , u > z 0 } . (50) \mathcal{D}_n:=\{\vartheta_n<+\infty\}\cap\left\{\inf_{u\in[\theta,\vartheta_n]}\widetilde{Y}_{1,u}>z_0\right\}. \tag*{(50)} D n := { ϑ n < + ∞ } ∩ { u ∈ [ θ , ϑ n ] inf Y 1 , u > z 0 } . (50)
Lemma 5.3. Suppose that the assumptions of Theorem 2 hold. For every T ∗ ∈ ( 0 , T ] T_\ast\in(0,T] T ∗ ∈ ( 0 , T ] and every integer n ≥ 1 n\ge1 n ≥ 1 , we have
P 0 ( D n ∩ { ϑ n ≤ θ + T ∗ 2 } | F θ ) > 0 a.s. on { n > 1 + sup u ∈ [ 0 , θ ] ∣ Z ‾ u ∣ } ∩ { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . \mathbb{P}_0\left(\mathcal{D}_n\cap\left\{\vartheta_n\le\theta+\frac{T_\ast}{2}\right\}\middle|\mathcal{F}_\theta\right)>0\quad\text{a.s. on}\quad\left\{n>1+\sup_{u\in[0,\theta]}|\overline{Z}_u|\right\}\cap\{\varepsilon_sX_\theta>x^\ast,sY_{1,\theta}>z_0\}. P 0 ( D n ∩ { ϑ n ≤ θ + 2 T ∗ } F θ ) > 0 a.s. on { n > 1 + u ∈ [ 0 , θ ] sup ∣ Z u ∣ } ∩ { ε s X θ > x ∗ , s Y 1 , θ > z 0 } .
Proof. We write ( X ~ , Y ~ ) = ( X ~ θ , Y ~ θ ) (\widetilde{X},\widetilde{Y})=(\widetilde{X}^{\theta},\widetilde{Y}^{\theta}) ( X , Y ) = ( X θ , Y θ ) and recall that ( X ~ θ , Y ~ θ ) = ( X θ , Y θ ) (\widetilde{X}_\theta,\widetilde{Y}_\theta)=(X_\theta,Y_\theta) ( X θ , Y θ ) = ( X θ , Y θ ) . We work conditionally on F θ \mathcal{F}_\theta F θ , on { n > 1 + sup u ∈ [ 0 , θ ] ∣ Z ‾ u ∣ } ∩ { ε s X θ > x ∗ , s Y 1 , θ > z 0 } \left\{n>1+\sup_{u\in[0,\theta]}|\overline{Z}_u|\right\}\cap\{\varepsilon_sX_\theta>x^\ast,sY_{1,\theta}>z_0\} { n > 1 + sup u ∈ [ 0 , θ ] ∣ Z u ∣ } ∩ { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . We use Lemma 5.2 to make Z ‾ \overline{Z} Z reach n n n while keeping ε s X ~ > x ∗ \varepsilon_s\widetilde{X}>x^\ast ε s X > x ∗ and s Y ~ 1 > z 0 s\widetilde{Y}_1>z_0 s Y 1 > z 0 . Continuity of F 1 , F 2 F_1,F_2 F 1 , F 2 gives δ ∈ ( 0 , 1 2 ] \delta\in(0,\frac{1}{2}] δ ∈ ( 0 , 2 1 ] such that s Y 1 , θ − 3 δ > z 0 sY_{1,\theta}-3\delta>z_0 s Y 1 , θ − 3 δ > z 0 and, for w ∈ [ Y 2 , θ − δ , Y 2 , θ + δ ] w\in[Y_{2,\theta}-\delta,Y_{2,\theta}+\delta] w ∈ [ Y 2 , θ − δ , Y 2 , θ + δ ] ,
ε s β ( θ ) − δ + ε s F 1 ( Y 1 , θ − 3 s δ ) + ε s F 2 ( w ) ≥ ε s X θ + x ∗ 2 . \varepsilon_s\beta(\theta)-\delta+\varepsilon_sF_1(Y_{1,\theta}-3s\delta)+\varepsilon_sF_2(w)\ge\frac{\varepsilon_sX_\theta+x^\ast}{2}. ε s β ( θ ) − δ + ε s F 1 ( Y 1 , θ − 3 sδ ) + ε s F 2 ( w ) ≥ 2 ε s X θ + x ∗ .
For h ∈ ( 0 , T ∗ 2 ] h\in(0,\frac{T_\ast}{2}] h ∈ ( 0 , 2 T ∗ ] , we set
τ = h ∧ inf { r ≥ 0 : Z ‾ θ + r ∉ ( Z ‾ θ − 2 δ , n ) or ∣ Y ~ 2 , θ + r − Y 2 , θ ∣ ≥ δ } . \tau=h\wedge\inf\left\{r\ge0:\overline{Z}_{\theta+r}\notin(\overline{Z}_\theta-2\delta,n)\ \text{or}\ |\widetilde{Y}_{2,\theta+r}-Y_{2,\theta}|\ge\delta\right\}. τ = h ∧ inf { r ≥ 0 : Z θ + r ∈ / ( Z θ − 2 δ , n ) or ∣ Y 2 , θ + r − Y 2 , θ ∣ ≥ δ } .
Since n > 1 + sup u ∈ [ 0 , θ ] ∣ Z ‾ u ∣ n>1+\sup_{u\in[0,\theta]}|\overline{Z}_u| n > 1 + sup u ∈ [ 0 , θ ] ∣ Z u ∣ and δ ≤ 1 2 \delta\le\frac12 δ ≤ 2 1 , the definition of τ \tau τ gives
sup u ∈ [ 0 , θ + τ ] ∣ Z ‾ u ∣ ≤ n , sup u ∈ [ 0 , θ + τ ] ∣ Y ~ 2 , u ∣ < 1 + sup u ∈ [ 0 , θ ] ∣ Y ~ 2 , u ∣ . \sup_{u\in[0,\theta+\tau]}|\overline{Z}_u|\le n,\qquad
\sup_{u\in[0,\theta+\tau]}|\widetilde{Y}_{2,u}|<1+\sup_{u\in[0,\theta]}|\widetilde{Y}_{2,u}|. u ∈ [ 0 , θ + τ ] sup ∣ Z u ∣ ≤ n , u ∈ [ 0 , θ + τ ] sup ∣ Y 2 , u ∣ < 1 + u ∈ [ 0 , θ ] sup ∣ Y 2 , u ∣.
Since β , G 1 \beta,G_1 β , G 1 are bounded on [ 0 , T ‾ ] [0,\overline{T}] [ 0 , T ] and F 1 , F 2 F_1,F_2 F 1 , F 2 are continuous, we have sup u ∈ [ 0 , θ + τ ] ∣ X ~ u ∣ ≤ C \sup_{u\in[0,\theta+\tau]}|\widetilde{X}_u|\le C sup u ∈ [ 0 , θ + τ ] ∣ X u ∣ ≤ C , where C C C does not depend on h h h . Indeed, Y ~ 1 = G 1 + s Z ‾ \widetilde{Y}_1=G_1+s\overline{Z} Y 1 = G 1 + s Z and the bounds on Z ‾ \overline{Z} Z and Y ~ 2 \widetilde{Y}_2 Y 2 give
sup u ∈ [ 0 , θ + τ ] ∣ X ~ u ∣ ≤ ∥ β ∥ ∞ , [ 0 , T ‾ ] + max ∣ y ∣ ≤ n + ∥ G 1 ∥ ∞ , [ 0 , T ‾ ] ∣ F 1 ( y ) ∣ + max ∣ w ∣ ≤ 1 + sup v ∈ [ 0 , θ ] ∣ Y ~ 2 , v ∣ ∣ F 2 ( w ) ∣ . \sup_{u\in[0,\theta+\tau]}|\widetilde{X}_u|
\le\|\beta\|_{\infty,[0,\overline{T}]}+
\max_{|y|\le n+\|G_1\|_{\infty,[0,\overline{T}]}}|F_1(y)|+
\max_{|w|\le1+\sup_{v\in[0,\theta]}|\widetilde{Y}_{2,v}|}|F_2(w)|. u ∈ [ 0 , θ + τ ] sup ∣ X u ∣ ≤ ∥ β ∥ ∞ , [ 0 , T ] + ∣ y ∣ ≤ n + ∥ G 1 ∥ ∞ , [ 0 , T ] max ∣ F 1 ( y ) ∣ + ∣ w ∣ ≤ 1 + s u p v ∈ [ 0 , θ ] ∣ Y 2 , v ∣ max ∣ F 2 ( w ) ∣.
We choose h h h small enough that
sup r ∈ [ 0 , h ] ( ∣ β ( θ + r ) − β ( θ ) ∣ ∨ ∣ G 1 ( θ + r ) − G 1 ( θ ) ∣ ) ≤ δ \sup_{r\in[0,h]}\left(|\beta(\theta+r)-\beta(\theta)|\vee|G_1(\theta+r)-G_1(\theta)|\right)\le\delta r ∈ [ 0 , h ] sup ( ∣ β ( θ + r ) − β ( θ ) ∣ ∨ ∣ G 1 ( θ + r ) − G 1 ( θ ) ∣ ) ≤ δ
and
sup r ∈ [ 0 , h ] ∣ G 2 ( θ + r ) − G 2 ( θ ) ∣ + C T ‾ ( 1 + C γ 2 ) h δ 2 < δ . \sup_{r\in[0,h]}|G_2(\theta+r)-G_2(\theta)|+C_{\overline{T}}(1+C^{\gamma_2})h^{\delta_2}<\delta. r ∈ [ 0 , h ] sup ∣ G 2 ( θ + r ) − G 2 ( θ ) ∣ + C T ( 1 + C γ 2 ) h δ 2 < δ .
The increment bound on Y ~ 2 \widetilde{Y}_2 Y 2 in (48) gives
∀ r ∈ [ 0 , τ ] , ∣ Y ~ 2 , θ + r − Y 2 , θ ∣ ≤ ∣ G 2 ( θ + r ) − G 2 ( θ ) ∣ + C T ‾ ( 1 + C γ 2 ) r δ 2 < δ . \forall r\in[0,\tau],\qquad
|\widetilde{Y}_{2,\theta+r}-Y_{2,\theta}|
\le|G_2(\theta+r)-G_2(\theta)|+C_{\overline{T}}(1+C^{\gamma_2})r^{\delta_2}<\delta. ∀ r ∈ [ 0 , τ ] , ∣ Y 2 , θ + r − Y 2 , θ ∣ ≤ ∣ G 2 ( θ + r ) − G 2 ( θ ) ∣ + C T ( 1 + C γ 2 ) r δ 2 < δ .
Thus, before h h h , only Z ˉ \bar{Z} Z ˉ can cause the stop at τ \tau τ . By the definition of τ \tau τ , the increment bound on G 1 G_1 G 1 , and monotonicity of z ↦ ε s F 1 ( s z ) z \mapsto\varepsilon_s F_1(sz) z ↦ ε s F 1 ( sz ) on [ z 0 , + ∞ ) [z_0,+\infty) [ z 0 , + ∞ ) , we have
∀ r ∈ [ 0 , τ ] , s Y ~ 1 , θ + r = s Y 1 , θ + Z ˉ θ + r − Z ˉ θ + s ( G 1 ( θ + r ) − G 1 ( θ ) ) ≥ s Y 1 , θ − 2 δ − δ = s Y 1 , θ − 3 δ > z 0 , \begin{aligned}
\forall r \in[0,\tau],\quad\widetilde{sY}_{1,\theta+r} &= sY_{1,\theta}+\bar{Z}_{\theta+r}-\bar{Z}_{\theta}+s\bigl(G_1(\theta+r)-G_1(\theta)\bigr)\\
&\ge sY_{1,\theta}-2\delta-\delta=sY_{1,\theta}-3\delta>z_0,
\end{aligned} ∀ r ∈ [ 0 , τ ] , s Y 1 , θ + r = s Y 1 , θ + Z ˉ θ + r − Z ˉ θ + s ( G 1 ( θ + r ) − G 1 ( θ ) ) ≥ s Y 1 , θ − 2 δ − δ = s Y 1 , θ − 3 δ > z 0 ,
ε s X ~ θ + r ≥ ε s β ( θ ) − δ + ε s F 1 ( Y 1 , θ − 3 s δ ) + ε s F 2 ( Y ~ 2 , θ + r ) ≥ ε s X θ + x ∗ 2 . (51) \begin{aligned}
\varepsilon_s\widetilde{X}_{\theta+r}
&\ge\varepsilon_s\beta(\theta)-\delta+\varepsilon_sF_1(Y_{1,\theta}-3s\delta)+\varepsilon_sF_2(\widetilde{Y}_{2,\theta+r})\\
&\ge\frac{\varepsilon_sX_\theta+x^*}{2}. \tag*{(51)}
\end{aligned} ε s X θ + r ≥ ε s β ( θ ) − δ + ε s F 1 ( Y 1 , θ − 3 sδ ) + ε s F 2 ( Y 2 , θ + r ) ≥ 2 ε s X θ + x ∗ . (51)
Since Γ 1 , σ \Gamma_1,\sigma Γ 1 , σ are bounded on [ 0 , T ‾ ] × [ − C , C ] [0,\overline{T}]\times[-C,C] [ 0 , T ] × [ − C , C ] , we may increase C C C so that ∣ σ ( θ + r , X ~ θ + r ) ∣ ≤ C |\sigma(\theta+r,\widetilde{X}_{\theta+r})|\le C ∣ σ ( θ + r , X θ + r ) ∣ ≤ C for r ∈ [ 0 , τ ] r\in[0,\tau] r ∈ [ 0 , τ ] . The drift lower bound (23) of Assumption 5.1 and the lower bound on ε s X ~ \varepsilon_s\widetilde{X} ε s X in (51) give
∀ r ∈ [ 0 , τ ] , C ∣ Γ 1 ( θ + r , X ~ θ + r ) ∣ ≥ s ρ Γ 1 ( θ + r , X ~ θ + r ) σ ( θ + r , X ~ θ + r ) ≥ c T ‾ ( ε s X θ + x ∗ 2 ) χ > 0. \forall r\in[0,\tau],\quad C|\Gamma_1(\theta+r,\widetilde{X}_{\theta+r})|\ge s\rho\Gamma_1(\theta+r,\widetilde{X}_{\theta+r})\sigma(\theta+r,\widetilde{X}_{\theta+r})\ge c_{\overline{T}}\left(\frac{\varepsilon_sX_\theta+x^*}{2}\right)^\chi>0. ∀ r ∈ [ 0 , τ ] , C ∣ Γ 1 ( θ + r , X θ + r ) ∣ ≥ s ρ Γ 1 ( θ + r , X θ + r ) σ ( θ + r , X θ + r ) ≥ c T ( 2 ε s X θ + x ∗ ) χ > 0.
Since K 1 ( 0 ) > 0 K_1(0)>0 K 1 ( 0 ) > 0 , the last lower bound and boundedness of Γ 1 \Gamma_1 Γ 1 give constants 0 < c ∗ ≤ C ∗ < ∞ 0<c_*\le C_*<\infty 0 < c ∗ ≤ C ∗ < ∞ , independent of h h h , such that
∀ r ∈ [ 0 , τ ] , c ∗ ≤ K 1 ( 0 ) ∣ Γ 1 ( θ + r , X ~ θ + r ) ∣ ≤ C ∗ . (52) \forall r\in[0,\tau],\quad c_*\le K_1(0)|\Gamma_1(\theta+r,\widetilde{X}_{\theta+r})|\le C_*. \tag*{(52)} ∀ r ∈ [ 0 , τ ] , c ∗ ≤ K 1 ( 0 ) ∣ Γ 1 ( θ + r , X θ + r ) ∣ ≤ C ∗ . (52)
The identity in (5.2) and the bound in (46) of Lemma 5.1 also bound the drift of Z ˉ θ + ⋅ \bar{Z}_{\theta+\cdot} Z ˉ θ +⋅ on [ 0 , τ ] [0,\tau] [ 0 , τ ] . Lemma 5.2, with
∀ r ∈ [ 0 , h ] , p ( r ) = r h ( n − Z ˉ θ + 2 δ ) , \forall r\in[0,h],\quad p(r)=\frac{r}{h}(n-\bar{Z}_{\theta}+2\delta), ∀ r ∈ [ 0 , h ] , p ( r ) = h r ( n − Z ˉ θ + 2 δ ) ,
gives
P 0 ( sup r ∈ [ 0 , τ ] ∣ Z ˉ θ + r − Z ˉ θ − p ( r ) ∣ < δ ∣ F θ ) > 0. \mathbb{P}_0\left(\left.\sup_{r\in[0,\tau]}|\bar{Z}_{\theta+r}-\bar{Z}_{\theta}-p(r)|<\delta\ \right|\mathcal{F}_\theta\right)>0. P 0 ( r ∈ [ 0 , τ ] sup ∣ Z ˉ θ + r − Z ˉ θ − p ( r ) ∣ < δ F θ ) > 0.
Here we have p ≥ 0 p\ge0 p ≥ 0 and Z ˉ θ + p ( h ) = n + 2 δ \bar{Z}_{\theta}+p(h)=n+2\delta Z ˉ θ + p ( h ) = n + 2 δ . On
{ sup r ∈ [ 0 , τ ] ∣ Z ˉ θ + r − Z ˉ θ − p ( r ) ∣ < δ } , \left\{\sup_{r\in[0,\tau]}|\bar{Z}_{\theta+r}-\bar{Z}_{\theta}-p(r)|<\delta\right\}, { r ∈ [ 0 , τ ] sup ∣ Z ˉ θ + r − Z ˉ θ − p ( r ) ∣ < δ } ,
we have
∀ r ∈ [ 0 , τ ] , Z ˉ θ + r > Z ˉ θ + p ( r ) − δ ≥ Z ˉ θ − δ > Z ˉ θ − 2 δ . \forall r\in[0,\tau],\quad\bar{Z}_{\theta+r}>\bar{Z}_{\theta}+p(r)-\delta\ge\bar{Z}_{\theta}-\delta>\bar{Z}_{\theta}-2\delta. ∀ r ∈ [ 0 , τ ] , Z ˉ θ + r > Z ˉ θ + p ( r ) − δ ≥ Z ˉ θ − δ > Z ˉ θ − 2 δ .
Still on this event, ϑ n > θ + τ \vartheta_n>\theta+\tau ϑ n > θ + τ would imply τ = h \tau=h τ = h and Z ˉ θ + h > Z ˉ θ + p ( h ) − δ = n + δ > n \bar{Z}_{\theta+h}>\bar{Z}_{\theta}+p(h)-\delta=n+\delta>n Z ˉ θ + h > Z ˉ θ + p ( h ) − δ = n + δ > n . This contradicts the definition of ϑ n \vartheta_n ϑ n . Therefore, we have ϑ n = θ + τ ≤ θ + T ‾ 2 \vartheta_n=\theta+\tau\le\theta+\frac{\overline{T}}{2} ϑ n = θ + τ ≤ θ + 2 T . The bound on s Y ~ 1 \widetilde{sY}_1 s Y 1 in (51) gives D n D_n D n . □ \square □
For a constant c 0 > 0 c_0>0 c 0 > 0 , chosen in Lemma 5.4 below, we set
h n : = c 0 n 1 − f 1 , s χ , τ n s t e p : = h n ∧ inf { r ≥ 0 : Z ˉ ϑ n + r ∉ ( n 2 , 2 n ) } . (53) h_n:=c_0n^{1-f_{1,s}\chi},\qquad\tau_n^{\mathrm{step}}:=h_n\wedge\inf\left\{r\ge0:\bar{Z}_{\vartheta_n+r}\notin\left(\frac{n}{2},2n\right)\right\}. \tag*{(53)} h n := c 0 n 1 − f 1 , s χ , τ n step := h n ∧ inf { r ≥ 0 : Z ˉ ϑ n + r ∈ / ( 2 n , 2 n ) } . (53)
We set τ n s t e p : = 0 \tau_n^{\mathrm{step}}:=0 τ n step := 0 on { ϑ n = ∞ } \{\vartheta_n=\infty\} { ϑ n = ∞ } . Events involving ( X ~ , Y ~ ) (\widetilde{X},\widetilde{Y}) ( X , Y ) at ϑ n \vartheta_n ϑ n include the condition ϑ n < ∞ \vartheta_n<\infty ϑ n < ∞ . The times used as Z ˉ \bar{Z} Z ˉ moves from each level n n n to 2 n 2n 2 n are listed in Table 7 .
Table 7. Times used in the explosion proof.
Lemma 5.4. Suppose that the assumptions of Theorem 2 hold. There exist deterministic constants T ∗ ∈ ( 0 , T ] T_* \in(0,T] T ∗ ∈ ( 0 , T ] and c 0 , C > 0 c_0,C>0 c 0 , C > 0 such that, for some a.s. finite F θ \mathcal{F}_{\theta} F θ -measurable integer N θ N_{\theta} N θ and every integer n ≥ 1 n\geq1 n ≥ 1 ,
P 0 ( D 2 n ∩ { ϑ 2 n ≤ ϑ n + h n } ∣ F ϑ n ) ≥ 1 − C h n (54) \mathbb{P}_0\left(\mathcal{D}_{2n}\cap\{\vartheta_{2n}\leq\vartheta_n+h_n\}\mid\mathcal{F}_{\vartheta_n}\right)\geq1-Ch_n \tag*{(54)} P 0 ( D 2 n ∩ { ϑ 2 n ≤ ϑ n + h n } ∣ F ϑ n ) ≥ 1 − C h n (54) a.s. on { n ≥ N θ } ∩ D n ∩ { ϑ n + h n ≤ θ + T ∗ } \{n\geq N_{\theta}\}\cap\mathcal{D}_n\cap\{\vartheta_n+h_n\leq\theta+T_*\} { n ≥ N θ } ∩ D n ∩ { ϑ n + h n ≤ θ + T ∗ } , where D n \mathcal{D}_n D n and h n h_n h n are defined in (50) and (53) , respectively.
Proof. We work on { n ≥ N θ } ∩ D n ∩ { ϑ n + h n ≤ θ + T ∗ } \{n\geq N_{\theta}\}\cap\mathcal{D}_n\cap\{\vartheta_n+h_n\leq\theta+T_*\} { n ≥ N θ } ∩ D n ∩ { ϑ n + h n ≤ θ + T ∗ } , with T ∗ ∈ ( 0 , T ] T_*\in(0,T] T ∗ ∈ ( 0 , T ] chosen below. The integer N θ N_{\theta} N θ is chosen large enough for all the bounds below; it depends on the finite quantities B B B from (49) , ∥ G 1 ∥ ∞ , [ 0 , T ‾ ] \lVert G_1\rVert_{\infty,[0,\overline{T}]} ∥ G 1 ∥ ∞ , [ 0 , T ] , and sup u ∈ [ 0 , θ ] ∣ Z ‾ u ∣ \sup_{u\in[0,\theta]}|\overline{Z}_u| sup u ∈ [ 0 , θ ] ∣ Z u ∣ . We first stop at τ ~ R X \widetilde{\tau}^{X}_R τ R X and prove the estimates with constants independent of R R R . By the definition of τ n s t e p \tau_n^{\mathrm{step}} τ n step in (53) , we have
∀ r ∈ [ 0 , τ n s t e p ] , n 2 ≤ Z ‾ ϑ n + r ≤ 2 n . (55) \forall r\in[0,\tau_n^{\mathrm{step}}],\qquad\frac{n}{2}\leq\overline{Z}_{\vartheta_n+r}\leq2n. \tag*{(55)} ∀ r ∈ [ 0 , τ n step ] , 2 n ≤ Z ϑ n + r ≤ 2 n . (55)
On D n \mathcal{D}_n D n , the definition of ϑ n \vartheta_n ϑ n gives
∀ u ∈ [ θ , ϑ n ] , z 0 − ∥ G 1 ∥ ∞ , [ 0 , T ] < Z ‾ u ≤ n . (56) \forall u\in[\theta,\vartheta_n],\qquad z_0-\lVert G_1\rVert_{\infty,[0,T]}<\overline{Z}_u\leq n. \tag*{(56)} ∀ u ∈ [ θ , ϑ n ] , z 0 − ∥ G 1 ∥ ∞ , [ 0 , T ] < Z u ≤ n . (56)
Since Y ~ 1 = Z ‾ + s G 1 \widetilde{Y}_1=\overline{Z}+sG_1 Y 1 = Z + s G 1 , the growth bounds on F 1 F_1 F 1 in Assumption 1.2, applied with z = s Y ~ 1 , u z=s\widetilde{Y}_{1,u} z = s Y 1 , u , and boundedness of G 1 G_1 G 1 give, for large n n n ,
inf r ∈ [ 0 , τ n s t e p ] ε s F 1 ( Y ~ 1 , ϑ n + r ) ≥ c 1 n f 1 , s , sup u ∈ [ θ , ϑ n + τ n s t e p ] ∣ F 1 ( Y ~ 1 , u ) ∣ ≤ C 1 n f 1 , s . \inf_{r\in[0,\tau_n^{\mathrm{step}}]}\varepsilon_sF_1(\widetilde{Y}_{1,\vartheta_n+r})\geq c_1n^{f_{1,s}},\qquad
\sup_{u\in[\theta,\vartheta_n+\tau_n^{\mathrm{step}}]}|F_1(\widetilde{Y}_{1,u})|\leq C_1n^{f_{1,s}}. r ∈ [ 0 , τ n step ] inf ε s F 1 ( Y 1 , ϑ n + r ) ≥ c 1 n f 1 , s , u ∈ [ θ , ϑ n + τ n step ] sup ∣ F 1 ( Y 1 , u ) ∣ ≤ C 1 n f 1 , s .
After increasing B B B to absorb ∥ β ∥ ∞ , [ 0 , T ‾ ] \lVert\beta\rVert_{\infty,[0,\overline{T}]} ∥ β ∥ ∞ , [ 0 , T ] and sup u ∈ [ 0 , θ ] ∣ X u ∣ \sup_{u\in[0,\theta]}|X_u| sup u ∈ [ 0 , θ ] ∣ X u ∣ , the bound on F 2 ( Y ~ 2 ) F_2(\widetilde{Y}_2) F 2 ( Y 2 ) in (49) gives, for t ∈ [ ϑ n , ϑ n + τ n s t e p ] t\in[\vartheta_n,\vartheta_n+\tau_n^{\mathrm{step}}] t ∈ [ ϑ n , ϑ n + τ n step ] ,
sup u ∈ [ 0 , t ] ∣ X ~ u ∣ ≤ B + C 1 n f 1 , s + C T ∗ δ 2 f 2 sup u ∈ [ 0 , t ] ∣ X ~ u ∣ . \sup_{u\in[0,t]}|\widetilde{X}_u|\leq B+C_1n^{f_{1,s}}+CT_*^{\delta_2f_2}\sup_{u\in[0,t]}|\widetilde{X}_u|. u ∈ [ 0 , t ] sup ∣ X u ∣ ≤ B + C 1 n f 1 , s + C T ∗ δ 2 f 2 u ∈ [ 0 , t ] sup ∣ X u ∣.
We choose T ∗ ∈ ( 0 , T ] T_*\in(0,T] T ∗ ∈ ( 0 , T ] such that
C T ∗ δ 2 f 2 ≤ min ( 1 2 , c 1 4 C 1 ) . CT_*^{\delta_2f_2}\leq\min\left(\frac{1}{2},\frac{c_1}{4C_1}\right). C T ∗ δ 2 f 2 ≤ min ( 2 1 , 4 C 1 c 1 ) .
It follows that
sup u ∈ [ 0 , t ] ∣ X ~ u ∣ ≤ 2 B + 2 C 1 n f 1 , s . \sup_{u\in[0,t]}|\widetilde{X}_u|\leq2B+2C_1n^{f_{1,s}}. u ∈ [ 0 , t ] sup ∣ X u ∣ ≤ 2 B + 2 C 1 n f 1 , s .
By estimate (49) , we have
∣ F 2 ( Y ~ 2 , t ) ∣ ≤ B + C T ∗ δ 2 f 2 ( 2 B + 2 C 1 n f 1 , s ) ≤ 2 B + c 1 2 n f 1 , s . |F_2(\widetilde{Y}_{2,t})|\leq B+CT_*^{\delta_2f_2}(2B+2C_1n^{f_{1,s}})\leq2B+\frac{c_1}{2}n^{f_{1,s}}. ∣ F 2 ( Y 2 , t ) ∣ ≤ B + C T ∗ δ 2 f 2 ( 2 B + 2 C 1 n f 1 , s ) ≤ 2 B + 2 c 1 n f 1 , s .
For all sufficiently large n n n , we have
ε s X ~ t ≥ ε s F 1 ( Y ~ 1 , t ) − ∣ F 2 ( Y ~ 2 , t ) ∣ − ∣ β ( t ) ∣ ≥ c 1 2 n f 1 , s − 2 B − ∥ β ∥ ∞ , [ 0 , T ‾ ] ≥ c 1 4 n f 1 , s . \varepsilon_s\widetilde{X}_t\geq\varepsilon_sF_1(\widetilde{Y}_{1,t})-|F_2(\widetilde{Y}_{2,t})|-|\beta(t)|\geq\frac{c_1}{2}n^{f_{1,s}}-2B-\lVert\beta\rVert_{\infty,[0,\overline{T}]}\geq\frac{c_1}{4}n^{f_{1,s}}. ε s X t ≥ ε s F 1 ( Y 1 , t ) − ∣ F 2 ( Y 2 , t ) ∣ − ∣ β ( t ) ∣ ≥ 2 c 1 n f 1 , s − 2 B − ∥ β ∥ ∞ , [ 0 , T ] ≥ 4 c 1 n f 1 , s .
For sufficiently large n n n , the upper bound on sup u ∈ [ 0 , t ] ∣ X ~ u ∣ \sup_{u\in[0,t]}|\widetilde{X}_u| sup u ∈ [ 0 , t ] ∣ X u ∣ and the lower bound on ε s X ~ t \varepsilon_s\widetilde{X}_t ε s X t imply
inf r ∈ [ 0 , τ n s t e p ] ε s X ~ ϑ n + r ≥ c n f 1 , s , (57) \inf_{r\in[0,\tau_n^{\mathrm{step}}]}\varepsilon_s\widetilde{X}_{\vartheta_n+r}\geq cn^{f_{1,s}}, \tag*{(57)} r ∈ [ 0 , τ n step ] inf ε s X ϑ n + r ≥ c n f 1 , s , (57)
sup u ∈ [ 0 , ϑ n + τ n s t e p ] ∣ X ~ u ∣ ≤ C n f 1 , s . (58) \sup_{u\in[0,\vartheta_n+\tau_n^{\mathrm{step}}]}|\widetilde{X}_u|\leq Cn^{f_{1,s}}. \tag*{(58)} u ∈ [ 0 , ϑ n + τ n step ] sup ∣ X u ∣ ≤ C n f 1 , s . (58)
For R > C n f 1 , s R>Cn^{f_{1,s}} R > C n f 1 , s , we can remove the extra stop at τ ~ R X \widetilde{\tau}^{X}_R τ R X . By the lower bound on ε s X ~ \varepsilon_s\widetilde{X} ε s X in (57) and the drift lower bound (23) of Assumption 5.1, we have
∀ u ∈ [ ϑ n , ϑ n + τ n s t e p ] , s ρ K 1 ( 0 ) Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) ≥ 2 c n f 1 , s χ , \forall u\in[\vartheta_n,\vartheta_n+\tau_n^{\mathrm{step}}],\qquad s\rho K_1(0)\Gamma_1(u,\widetilde{X}_u)\sigma(u,\widetilde{X}_u)\geq2cn^{f_{1,s}}\chi, ∀ u ∈ [ ϑ n , ϑ n + τ n step ] , s ρ K 1 ( 0 ) Γ 1 ( u , X u ) σ ( u , X u ) ≥ 2 c n f 1 , s χ ,
for some c > 0 c > 0 c > 0 . Since Z ˉ \bar{Z} Z ˉ is bounded on [ 0 , θ ] [0,\theta] [ 0 , θ ] , the bounds (56) and (55) give sup u ∈ [ 0 , ϑ n + τ n s t e p ] ∣ Z ˉ u ∣ ≤ 2 n \sup_{u\in[0,\vartheta_n+\tau_n^{\mathrm{step}}]}|\bar{Z}_u| \le2n sup u ∈ [ 0 , ϑ n + τ n step ] ∣ Z ˉ u ∣ ≤ 2 n for large n n n . We set
∀ r ∈ [ 0 , h n ] , M r : = s K 1 ( 0 ) ∫ ϑ n ϑ n + ( r ∧ τ n s t e p ) Γ 1 ( u , X ~ u ) d W u . \forall r \in[0,h_n],\qquad M_r := sK_1(0)\int_{\vartheta_n}^{\vartheta_n+(r\wedge\tau_n^{\mathrm{step}})}\Gamma_1(u,\widetilde{X}_u)\,\mathrm{d}W_u. ∀ r ∈ [ 0 , h n ] , M r := s K 1 ( 0 ) ∫ ϑ n ϑ n + ( r ∧ τ n step ) Γ 1 ( u , X u ) d W u .
We write ⟨ M ⟩ \langle M\rangle ⟨ M ⟩ for the predictable quadratic variation of M M M . Subtracting (5.2) of Lemma 5.1 at ϑ n \vartheta_n ϑ n from the same identity at ϑ n + r \vartheta_n+r ϑ n + r , and using Z ˉ ϑ n = n \bar{Z}_{\vartheta_n}=n Z ˉ ϑ n = n , gives, by (46) of the same lemma, with R > C n f 1 , s R>Cn^{f_{1,s}} R > C n f 1 , s from (58) , for large n n n , since f 1 , s χ > 1 f_{1,s}\chi>1 f 1 , s χ > 1 ,
∀ r ∈ [ 0 , τ n s t e p ] , Z ˉ ϑ n + r = n + M r + s ρ K 1 ( 0 ) ∫ ϑ n ϑ n + r Γ 1 ( u , X ~ u ) σ ( u , X ~ u ) d u \forall r\in[0,\tau_n^{\mathrm{step}}],\qquad\bar{Z}_{\vartheta_n+r}=n+M_r+s\rho K_1(0)\int_{\vartheta_n}^{\vartheta_n+r}\Gamma_1(u,\widetilde{X}_u)\sigma(u,\widetilde{X}_u)\,\mathrm{d}u ∀ r ∈ [ 0 , τ n step ] , Z ˉ ϑ n + r = n + M r + s ρ K 1 ( 0 ) ∫ ϑ n ϑ n + r Γ 1 ( u , X u ) σ ( u , X u ) d u
+ ∫ ϑ n ϑ n + r ( r m e m ( 0 ) Z ˉ u + ( d r m e m ∗ Z ˉ ) ( u ) ) d u ≥ n + ( 2 c n f 1 , s χ − C n ) r + M r ≥ n + c n f 1 , s χ r + M r . (59) \begin{aligned}
\qquad\qquad+\int_{\vartheta_n}^{\vartheta_n+r}\left(r_{\mathrm{mem}}(0)\bar{Z}_u+(\mathrm{d}r_{\mathrm{mem}}*\bar{Z})(u)\right)\,\mathrm{d}u \\
&\ge n+(2cn^{f_{1,s}\chi}-Cn)r+M_r \\
&\ge n+cn^{f_{1,s}\chi}r+M_r. \tag*{(59)}
\end{aligned} + ∫ ϑ n ϑ n + r ( r mem ( 0 ) Z ˉ u + ( d r mem ∗ Z ˉ ) ( u ) ) d u ≥ n + ( 2 c n f 1 , s χ − C n ) r + M r ≥ n + c n f 1 , s χ r + M r . (59)
By the polynomial growth bound on Γ 1 \Gamma_1 Γ 1 in Assumption 1.2 and γ 1 f 1 , s ≤ γ 1 f 1 ≤ 1 \gamma_1f_{1,s}\le\gamma_1f_1\le1 γ 1 f 1 , s ≤ γ 1 f 1 ≤ 1 , we have
⟨ M ⟩ h n = K 1 ( 0 ) 2 ∫ ϑ n ϑ n + τ n s t e p ∣ Γ 1 ( u , X ~ u ) ∣ 2 d u ≤ C ( 1 + sup u ∈ [ 0 , ϑ n + τ n s t e p ] ∣ X ~ u ∣ 2 γ 1 ) h n ≤ C n 2 γ 1 f 1 , s h n ≤ C n 2 h n . (60) \begin{aligned}
\langle M\rangle_{h_n}=K_1(0)^2\int_{\vartheta_n}^{\vartheta_n+\tau_n^{\mathrm{step}}}|\Gamma_1(u,\widetilde{X}_u)|^2\,\mathrm{d}u \\
&\le C\left(1+\sup_{u\in[0,\vartheta_n+\tau_n^{\mathrm{step}}]}|\widetilde{X}_u|^{2\gamma_1}\right)h_n\le Cn^{2\gamma_1f_{1,s}}h_n\le Cn^2h_n. \tag*{(60)}
\end{aligned} ⟨ M ⟩ h n = K 1 ( 0 ) 2 ∫ ϑ n ϑ n + τ n step ∣ Γ 1 ( u , X u ) ∣ 2 d u ≤ C ( 1 + u ∈ [ 0 , ϑ n + τ n step ] sup ∣ X u ∣ 2 γ 1 ) h n ≤ C n 2 γ 1 f 1 , s h n ≤ C n 2 h n . (60)
For sufficiently large n n n , the identity s Y ~ 1 = Z ˉ + s G 1 s\widetilde{Y}_1=\bar{Z}+sG_1 s Y 1 = Z ˉ + s G 1 gives
inf r ∈ [ 0 , τ n s t e p ] Y ~ 1 , ϑ n + r ≥ n 2 − ∥ G 1 ∥ ∞ , [ 0 , T ‾ ] > z 0 , \inf_{r\in[0,\tau_n^{\mathrm{step}}]}\widetilde{Y}_{1,\vartheta_n+r}\ge\frac{n}{2}-\|G_1\|_{\infty,[0,\overline{T}]} > z_0, r ∈ [ 0 , τ n step ] inf Y 1 , ϑ n + r ≥ 2 n − ∥ G 1 ∥ ∞ , [ 0 , T ] > z 0 ,
so Z ˉ ϑ n + τ n s t e p = 2 n \bar{Z}_{\vartheta_n+\tau_n^{\mathrm{step}}}=2n Z ˉ ϑ n + τ n step = 2 n implies D 2 n D_{2n} D 2 n and ϑ 2 n = ϑ n + τ n s t e p ≤ ϑ n + h n \vartheta_{2n}=\vartheta_n+\tau_n^{\mathrm{step}}\le\vartheta_n+h_n ϑ 2 n = ϑ n + τ n step ≤ ϑ n + h n . In the definition of h n h_n h n in (53) , we choose c 0 : = 2 c c_0:=\frac{2}{c} c 0 := c 2 , where c > 0 c>0 c > 0 is the coefficient of n f 1 , s χ r n^{f_{1,s}\chi}r n f 1 , s χ r in the lower bound (59) . If Z ˉ ϑ n + τ n s t e p ≠ 2 n \bar{Z}_{\vartheta_n+\tau_n^{\mathrm{step}}}\ne2n Z ˉ ϑ n + τ n step = 2 n , the lower bound on Z ˉ \bar{Z} Z ˉ in (59) and c n f 1 , s χ h n = 2 n cn^{f_{1,s}\chi}h_n=2n c n f 1 , s χ h n = 2 n give
M τ n s t e p ≤ Z ˉ ϑ n + τ n s t e p − n − c n f 1 , s χ τ n s t e p ≤ { n 2 − n = − n 2 , Z ˉ ϑ n + τ n s t e p = n 2 , 2 n − n − c n f 1 , s χ h n = − n , τ n s t e p = h n . M_{\tau_n^{\mathrm{step}}}\le\bar{Z}_{\vartheta_n+\tau_n^{\mathrm{step}}}-n-cn^{f_{1,s}\chi}\tau_n^{\mathrm{step}}
\le
\begin{cases}
\frac{n}{2}-n=-\frac{n}{2}, & \bar{Z}_{\vartheta_n+\tau_n^{\mathrm{step}}}=\frac{n}{2},\\
2n-n-cn^{f_{1,s}\chi}h_n=-n, & \tau_n^{\mathrm{step}}=h_n.
\end{cases} M τ n step ≤ Z ˉ ϑ n + τ n step − n − c n f 1 , s χ τ n step ≤ { 2 n − n = − 2 n , 2 n − n − c n f 1 , s χ h n = − n , Z ˉ ϑ n + τ n step = 2 n , τ n step = h n .
Doob’s inequality applied to M 2 M^2 M 2 and the quadratic variation bound (60) give
P 0 ( inf r ∈ [ 0 , h n ] M r ≤ − n 2 | F ϑ n ) ≤ 4 n 2 E 0 [ ⟨ M ⟩ h n | F ϑ n ] ≤ C h n , \mathbb{P}_0\left(\inf_{r\in[0,h_n]}M_r\le-\frac{n}{2}\,\middle|\,\mathcal{F}_{\vartheta_n}\right)
\le\frac{4}{n^2}\mathbb{E}_0\left[\langle M\rangle_{h_n}\,\middle|\,\mathcal{F}_{\vartheta_n}\right]\le Ch_n, P 0 ( r ∈ [ 0 , h n ] inf M r ≤ − 2 n F ϑ n ) ≤ n 2 4 E 0 [ ⟨ M ⟩ h n ∣ F ϑ n ] ≤ C h n ,
which proves (54) of Lemma 5.4. □
Proof of Theorem 2(i). We use the constants T ∗ T_* T ∗ , c 0 c_0 c 0 , C C C and the integer N θ N_\theta N θ from Lemma 5.4, and the times h n h_n h n defined in (53) . There exists an a.s. finite F θ \mathcal{F}_\theta F θ -measurable integer n 0 ≥ N θ n_0\ge N_\theta n 0 ≥ N θ , such that n 0 > 1 + sup u ∈ [ 0 , θ ] ∣ Z ˉ u ∣ n_0>1+\sup_{u\in[0,\theta]}|\bar{Z}_u| n 0 > 1 + sup u ∈ [ 0 , θ ] ∣ Z ˉ u ∣ , and
∑ k = 0 ∞ h 2 k n 0 ≤ min ( T ∗ 2 , 1 2 C ) . (61) \sum_{k=0}^{\infty}h_{2^kn_0}\le\min\left(\frac{T_*}{2},\frac{1}{2C}\right). \tag*{(61)} k = 0 ∑ ∞ h 2 k n 0 ≤ min ( 2 T ∗ , 2 C 1 ) . (61)
Indeed, Z ˉ \bar{Z} Z ˉ is continuous on [ 0 , θ ] [0,\theta] [ 0 , θ ] and, since f 1 , s χ > 1 f_{1,s}\chi>1 f 1 , s χ > 1 , we have
∑ k = 0 ∞ h 2 k n 0 = c 0 n 0 1 − f 1 , s χ ∑ k = 0 ∞ ( 2 1 − f 1 , s χ ) k = c 0 n 0 1 − f 1 , s χ 1 − 2 1 − f 1 , s χ → n 0 → ∞ 0. \sum_{k=0}^{\infty}h_{2^kn_0}
=c_0n_0^{1-f_{1,s}\chi}\sum_{k=0}^{\infty}(2^{1-f_{1,s}\chi})^k
=\frac{c_0n_0^{1-f_{1,s}\chi}}{1-2^{1-f_{1,s}\chi}}
\xrightarrow[n_0\to\infty]{}0. k = 0 ∑ ∞ h 2 k n 0 = c 0 n 0 1 − f 1 , s χ k = 0 ∑ ∞ ( 2 1 − f 1 , s χ ) k = 1 − 2 1 − f 1 , s χ c 0 n 0 1 − f 1 , s χ n 0 → ∞ 0.
The events
H 0 : = D n 0 ∩ { ϑ n 0 ≤ θ + T ∗ 2 } , \mathcal{H}_0 := \mathcal{D}_{n_0} \cap\left\{\vartheta_{n_0} \le\theta+ \frac{T_*}{2}\right\}, H 0 := D n 0 ∩ { ϑ n 0 ≤ θ + 2 T ∗ } ,
∀ N ∈ N , H N + 1 : = H N ∩ D 2 N + 1 n 0 ∩ { ϑ 2 N n 0 ≤ ϑ 2 N n 0 + h 2 N n 0 } , \forall N \in\mathbb{N}, \qquad\mathcal{H}_{N+1} := \mathcal{H}_N \cap\mathcal{D}_{2^{N+1}n_0} \cap\left\{\vartheta_{2^N n_0} \le\vartheta_{2^N n_0} + h_{2^N n_0}\right\}, ∀ N ∈ N , H N + 1 := H N ∩ D 2 N + 1 n 0 ∩ { ϑ 2 N n 0 ≤ ϑ 2 N n 0 + h 2 N n 0 } ,
satisfy H N ∈ F ϑ 2 N n 0 \mathcal{H}_N \in\mathcal{F}_{\vartheta_{2^N n_0}} H N ∈ F ϑ 2 N n 0 . The conditional estimates at the random levels 2 N n 0 2^N n_0 2 N n 0 follow by partitioning over { n 0 = j } ∈ F θ \{n_0=j\}\in\mathcal{F}_\theta { n 0 = j } ∈ F θ , j ≥ 1 j\ge1 j ≥ 1 . By Lemma 5.3, we have
P 0 ( H 0 ∣ F θ ) > 0 a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . \mathbb{P}_0(\mathcal{H}_0\mid\mathcal{F}_\theta)>0 \quad\text{a.s. on } \{\varepsilon_sX_\theta>x^*,\ sY_{1,\theta}>z_0\}. P 0 ( H 0 ∣ F θ ) > 0 a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } .
On H N \mathcal{H}_N H N , we have
ϑ 2 N n 0 + h 2 N n 0 ≤ θ + T ∗ 2 + ∑ k = 0 N h 2 k n 0 ≤ θ + T ∗ . \vartheta_{2^Nn_0}+h_{2^Nn_0}\le\theta+\frac{T_*}{2}+\sum_{k=0}^{N}h_{2^kn_0}\le\theta+T_*. ϑ 2 N n 0 + h 2 N n 0 ≤ θ + 2 T ∗ + k = 0 ∑ N h 2 k n 0 ≤ θ + T ∗ .
By Lemma 5.4 and the tower property, for N ≥ 0 N\ge0 N ≥ 0 , we have
P 0 ( H N + 1 ∣ F θ ) = E 0 [ 1 H N P 0 ( H N + 1 ∣ F ϑ 2 N n 0 ) | F θ ] ≥ ( 1 − C h 2 N n 0 ) P 0 ( H N ∣ F θ ) . \mathbb{P}_0(\mathcal{H}_{N+1}\mid\mathcal{F}_\theta)=\mathbb{E}_0\left[\mathbf{1}_{\mathcal{H}_N}\mathbb{P}_0\left(\mathcal{H}_{N+1}\mid\mathcal{F}_{\vartheta_{2^Nn_0}}\right)\middle|\mathcal{F}_\theta\right]\ge(1-Ch_{2^Nn_0})\mathbb{P}_0(\mathcal{H}_N\mid\mathcal{F}_\theta). P 0 ( H N + 1 ∣ F θ ) = E 0 [ 1 H N P 0 ( H N + 1 ∣ F ϑ 2 N n 0 ) F θ ] ≥ ( 1 − C h 2 N n 0 ) P 0 ( H N ∣ F θ ) .
By induction, for N ≥ 1 N\ge1 N ≥ 1 , we have
P 0 ( H N ∣ F θ ) ≥ P 0 ( H 0 ∣ F θ ) ∏ k = 0 N − 1 ( 1 − C h 2 k n 0 ) ≥ P 0 ( H 0 ∣ F θ ) ( 1 − C ∑ k = 0 N − 1 h 2 k n 0 ) ≥ 1 2 P 0 ( H 0 ∣ F θ ) . \mathbb{P}_0(\mathcal{H}_N\mid\mathcal{F}_\theta)\ge\mathbb{P}_0(\mathcal{H}_0\mid\mathcal{F}_\theta)\prod_{k=0}^{N-1}(1-Ch_{2^kn_0})\ge\mathbb{P}_0(\mathcal{H}_0\mid\mathcal{F}_\theta)\left(1-C\sum_{k=0}^{N-1}h_{2^kn_0}\right)\ge\frac{1}{2}\mathbb{P}_0(\mathcal{H}_0\mid\mathcal{F}_\theta). P 0 ( H N ∣ F θ ) ≥ P 0 ( H 0 ∣ F θ ) k = 0 ∏ N − 1 ( 1 − C h 2 k n 0 ) ≥ P 0 ( H 0 ∣ F θ ) ( 1 − C k = 0 ∑ N − 1 h 2 k n 0 ) ≥ 2 1 P 0 ( H 0 ∣ F θ ) .
Since the events H N \mathcal{H}_N H N decrease, we have
P 0 ( ⋂ N ≥ 0 H N | F θ ) ≥ 1 2 P 0 ( H 0 ∣ F θ ) > 0 a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . \mathbb{P}_0\left(\bigcap_{N\ge0}\mathcal{H}_N\middle|\mathcal{F}_\theta\right)\ge\frac{1}{2}\mathbb{P}_0(\mathcal{H}_0\mid\mathcal{F}_\theta)>0 \quad\text{a.s. on } \{\varepsilon_sX_\theta>x^*,\ sY_{1,\theta}>z_0\}. P 0 ( N ≥ 0 ⋂ H N F θ ) ≥ 2 1 P 0 ( H 0 ∣ F θ ) > 0 a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } .
On ⋂ N ≥ 0 H N \bigcap_{N\ge0}\mathcal{H}_N ⋂ N ≥ 0 H N , the lower bound on ε s X ~ \varepsilon_s\widetilde{X} ε s X in (57) gives
ϑ 2 N n 0 ≤ θ + T ∗ ≤ θ + T , ε s X ~ ϑ 2 N n 0 ≥ c ( 2 N n 0 ) f 1 , s ⟶ N → ∞ + ∞ . \vartheta_{2^Nn_0}\le\theta+T_*\le\theta+T,\qquad\varepsilon_s\widetilde{X}_{\vartheta_{2^Nn_0}}\ge c(2^Nn_0)^{f_{1,s}}\underset{N\to\infty}{\longrightarrow}+\infty. ϑ 2 N n 0 ≤ θ + T ∗ ≤ θ + T , ε s X ϑ 2 N n 0 ≥ c ( 2 N n 0 ) f 1 , s N → ∞ ⟶ + ∞.
Recalling from Assumption 2.2 that
τ ~ ( θ , n ) X ~ θ : = inf { u ≥ 0 : ∣ X ~ θ + u ∣ ≥ n } , \widetilde{\tau}^{\widetilde{X}_\theta}_{(\theta,n)}:=\inf\{u\ge0:|\widetilde{X}_{\theta+u}|\ge n\}, τ ( θ , n ) X θ := inf { u ≥ 0 : ∣ X θ + u ∣ ≥ n } ,
the bound ∣ X ~ ∣ ≥ ε s X ~ |\widetilde{X}|\ge\varepsilon_s\widetilde{X} ∣ X ∣ ≥ ε s X gives
lim n → ∞ P 0 ( τ ~ ( θ , n ) X ~ θ ≤ T ∣ F θ ) ≥ P 0 ( ⋂ N ≥ 0 H N | F θ ) > 0 (62) \lim_{n\to\infty}\mathbb{P}_0\left(\widetilde{\tau}^{\widetilde{X}_\theta}_{(\theta,n)}\le T\mid\mathcal{F}_\theta\right)\ge\mathbb{P}_0\left(\bigcap_{N\ge0}\mathcal{H}_N\middle|\mathcal{F}_\theta\right)>0 \tag*{(62)} n → ∞ lim P 0 ( τ ( θ , n ) X θ ≤ T ∣ F θ ) ≥ P 0 ( N ≥ 0 ⋂ H N F θ ) > 0 (62)
a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } \{\varepsilon_sX_\theta>x^*,\ sY_{1,\theta}>z_0\} { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . The bound in (62) and Lemma 3.1 give, for every bounded stopping time θ \theta θ ,
E 0 [ S θ + T ∣ F θ ] ≤ S θ (63) \mathbb{E}_0[S_{\theta+T}\mid\mathcal{F}_\theta]\le S_\theta \tag*{(63)} E 0 [ S θ + T ∣ F θ ] ≤ S θ (63)
a.s. on { ε s X θ > x ∗ , s Y 1 , θ > z 0 } \{\varepsilon_sX_\theta>x^*,\ sY_{1,\theta}>z_0\} { ε s X θ > x ∗ , s Y 1 , θ > z 0 } . With θ = t 0 \theta=t_0 θ = t 0 and T = t − t 0 T=t-t_0 T = t − t 0 , the supermartingale property gives E 0 [ S t ] ≤ E 0 [ S t 0 ] \mathbb{E}_0[S_t]\le\mathbb{E}_0[S_{t_0}] E 0 [ S t ] ≤ E 0 [ S t 0 ] whenever P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 \mathbb{P}_0(\varepsilon_sX_{t_0}>x^*,\ sY_{1,t_0}>z_0)>0 P 0 ( ε s X t 0 > x ∗ , s Y 1 , t 0 > z 0 ) > 0 . Thus S S S is a strict local martingale on [ t 0 , t ] [t_0,t] [ t 0 , t ] .
Refinement
The strict conditional inequality in Theorem 2 (ii) concerns the expected asset price after τ ∗ \tau_* τ ∗ , given F τ ∗ \mathcal{F}_{\tau_*} F τ ∗ , on the event { τ ∗ < ∞ } \{\tau_* < \infty\} { τ ∗ < ∞ } .
Proof of Theorem 2(ii). We fix T > 0 T > 0 T > 0 and choose any N ∈ N ∗ N \in\mathbb{N}^* N ∈ N ∗ such that P 0 ( τ ∗ ≤ N ) > 0 \mathbb{P}_0(\tau_* \le N) > 0 P 0 ( τ ∗ ≤ N ) > 0 . Such an integer exists since P 0 ( τ ∗ ≤ N ) ↑ P 0 ( τ ∗ < ∞ ) > 0 \mathbb{P}_0(\tau_* \le N) \uparrow\mathbb{P}_0(\tau_* < \infty) > 0 P 0 ( τ ∗ ≤ N ) ↑ P 0 ( τ ∗ < ∞ ) > 0 as N → ∞ N \to\infty N → ∞ . The event { τ ∗ ≤ N } \{\tau_* \le N\} { τ ∗ ≤ N } belongs to F τ ∗ \mathcal{F}_{\tau_*} F τ ∗ . Let A ∈ F τ ∗ A \in\mathcal{F}_{\tau_*} A ∈ F τ ∗ satisfy A ⊂ { τ ∗ ≤ N } A \subset\{\tau_* \le N\} A ⊂ { τ ∗ ≤ N } . We have A ∈ F τ ∗ ∧ N A \in\mathcal{F}_{\tau_* \wedge N} A ∈ F τ ∗ ∧ N . We assume that P 0 ( A ) > 0 \mathbb{P}_0(A) > 0 P 0 ( A ) > 0 . By the definition of τ ∗ \tau_* τ ∗ in Assumption 5.2 and continuity of X , Y 1 X,Y_1 X , Y 1 , we have
A = ⋃ h ∈ Q ∩ ( 0 , T ) A ∩ { ε s X τ ∗ + h > x ∗ , s Y 1 , τ ∗ + h > z 0 } . A=\bigcup_{h\in\mathbb{Q}\cap(0,T)}A\cap\{\varepsilon_sX_{\tau_*+h}>x^*,\ sY_{1,\tau_*+h}>z_0\}. A = h ∈ Q ∩ ( 0 , T ) ⋃ A ∩ { ε s X τ ∗ + h > x ∗ , s Y 1 , τ ∗ + h > z 0 } .
This union is countable, so one of its events has positive probability. For the corresponding h h h , (63) applies at ( τ ∗ ∧ N ) + h (\tau_* \wedge N)+h ( τ ∗ ∧ N ) + h , over the interval of length T − h T-h T − h , on A ∩ { ε s X τ ∗ + h > x ∗ , s Y 1 , τ ∗ + h > z 0 } A\cap\{\varepsilon_sX_{\tau_*+h}>x^*,\ sY_{1,\tau_*+h}>z_0\} A ∩ { ε s X τ ∗ + h > x ∗ , s Y 1 , τ ∗ + h > z 0 } . Since τ ∗ ∧ N = τ ∗ \tau_* \wedge N=\tau_* τ ∗ ∧ N = τ ∗ on A A A , the tower property and optional sampling give
E 0 [ 1 A S τ ∗ + T ] = E 0 [ 1 A E 0 [ S ( τ ∗ ∧ N ) + T ∣ F ( τ ∗ ∧ N ) + h ] ] < E 0 [ 1 A S ( τ ∗ ∧ N ) + h ] = E 0 [ 1 A E 0 [ S ( τ ∗ ∧ N ) + h ∣ F τ ∗ ∧ N ] ] ≤ E 0 [ 1 A S τ ∗ ] . \begin{aligned}
\mathbb{E}_0[\mathbf{1}_A S_{\tau_*+T}]&=\mathbb{E}_0\left[\mathbf{1}_A\mathbb{E}_0\left[S_{(\tau_*\wedge N)+T}\mid\mathcal{F}_{(\tau_*\wedge N)+h}\right]\right]\\
&<\mathbb{E}_0[\mathbf{1}_A S_{(\tau_*\wedge N)+h}]\\
&=\mathbb{E}_0\left[\mathbf{1}_A\mathbb{E}_0\left[S_{(\tau_*\wedge N)+h}\mid\mathcal{F}_{\tau_*\wedge N}\right]\right]\\
&\le\mathbb{E}_0[\mathbf{1}_A S_{\tau_*}].
\end{aligned} E 0 [ 1 A S τ ∗ + T ] = E 0 [ 1 A E 0 [ S ( τ ∗ ∧ N ) + T ∣ F ( τ ∗ ∧ N ) + h ] ] < E 0 [ 1 A S ( τ ∗ ∧ N ) + h ] = E 0 [ 1 A E 0 [ S ( τ ∗ ∧ N ) + h ∣ F τ ∗ ∧ N ] ] ≤ E 0 [ 1 A S τ ∗ ] .
Since A A A is arbitrary, the characterization of conditional expectation gives the claimed strict inequality on { τ ∗ ≤ N } \{\tau_* \le N\} { τ ∗ ≤ N } . The union over N N N proves the result on { τ ∗ < ∞ } \{\tau_* < \infty\} { τ ∗ < ∞ } . □
Conclusion
We have studied the martingale property of stochastic and path-dependent Volterra volatility models in a common setting. A Girsanov change of measure reduces martingality of S S S to non-explosion of X ˉ θ \bar{X}^{\theta} X ˉ θ in (PDV$_\theta$). On the one hand, we prove martingality either from a one-sided bound on the state, as in Assumption 3.1, or from a one-sided growth bound on the Girsanov drift, as in {xref}Assumption 3.2. On the other hand, for smooth kernels satisfying {xref}Assumption 4.2, the growth and monotonicity conditions in {xref}Assumption 5.1 and the accessibility condition in {xref}Assumption 5.2` give strict local martingality.
Our results also give the known martingality results for lognormal stochastic volatility and affine Volterra-type models. They give new results for the Volterra–Guyon–Lekeufack model, upside-enhanced Volterra–Guyon–Lekeufack models, the polynomial PDV model, and quadratic Volterra–Heston-type models. They show how the sign of σ \sigma σ , the correlation parameter, and the growth of Γ 1 σ \Gamma_1\sigma Γ 1 σ affect whether the price process is a true or strict local martingale.
To extend the explosion criterion to rough kernels, we would need a proof that does not use the semimartingale representation of Z ˉ \bar{Z} Z ˉ .
A fractional Sobolev estimate
Lemma A.1. For r > 2 r > 2 r > 2 , there exists C r > 0 C_r > 0 C r > 0 such that every real-valued process φ \varphi φ with continuous paths on [ 0 , T ] [0,T] [ 0 , T ] satisfies, for t ∈ ( 0 , T ] t \in(0,T] t ∈ ( 0 , T ] ,
E 0 [ sup u ∈ [ 0 , t ] ∣ φ u − φ 0 ∣ r ] ≤ C r t ∫ 0 t ∫ 0 t E 0 [ ∣ φ v − φ u ∣ r ] ∣ v − u ∣ 3 d u d v . (64) \mathbb{E}_0\left[\sup_{u\in[0,t]}|\varphi_u-\varphi_0|^r\right]\le C_rt\int_0^t\int_0^t\frac{\mathbb{E}_0[|\varphi_v-\varphi_u|^r]}{|v-u|^3}\,du\,dv. \tag*{(64)} E 0 [ u ∈ [ 0 , t ] sup ∣ φ u − φ 0 ∣ r ] ≤ C r t ∫ 0 t ∫ 0 t ∣ v − u ∣ 3 E 0 [ ∣ φ v − φ u ∣ r ] d u d v . (64)
This is a consequence of Corollary A.2 of [19 ] , with q = r q=r q = r and α = 2 r \alpha=\frac{2}{r} α = r 2 .
Proof of Lemma 5.1
Under the kernel regularity in Assumption 4.2, we apply [24 Chapter 2, Theorem 3.5] with
∀ t ∈ [ 0 , + ∞ ) , k ( t ) = K 1 ′ ( t ) K 1 ( 0 ) , f ( t ) = − K 1 ′ ( t ) K 1 ( 0 ) 2 . \forall t \in[0,+\infty),\qquad k(t)=\frac{K_1'(t)}{K_1(0)},\qquad f(t)=-\frac{K_1'(t)}{K_1(0)^2}. ∀ t ∈ [ 0 , + ∞ ) , k ( t ) = K 1 ( 0 ) K 1 ′ ( t ) , f ( t ) = − K 1 ( 0 ) 2 K 1 ′ ( t ) .
This gives a unique solution ℓ r e s ∈ C ( R + ) ∩ B V l o c ( R + ) \ell_{\mathrm{res}} \in C(\mathbb{R}_+) \cap BV_{\mathrm{loc}}(\mathbb{R}_+) ℓ res ∈ C ( R + ) ∩ B V loc ( R + ) to the Volterra equation below.
ℓ r e s + K 1 ′ K 1 ( 0 ) ∗ ℓ r e s = − K 1 ′ K 1 ( 0 ) 2 . (65) \ell_{\mathrm{res}}+\frac{K_1'}{K_1(0)} * \ell_{\mathrm{res}}=-\frac{K_1'}{K_1(0)^2}. \tag*{(65)} ℓ res + K 1 ( 0 ) K 1 ′ ∗ ℓ res = − K 1 ( 0 ) 2 K 1 ′ . (65)
We define R 1 \mathcal{R}_1 R 1 by
R 1 : = δ 0 K 1 ( 0 ) + ℓ r e s ( t ) d t . (66) \mathcal{R}_1:=\frac{\delta_0}{K_1(0)}+\ell_{\mathrm{res}}(t)\,dt. \tag*{(66)} R 1 := K 1 ( 0 ) δ 0 + ℓ res ( t ) d t . (66)
Since ( K 1 ∗ R 1 ) ( 0 ) = 1 (K_1 * \mathcal{R}_1)(0)=1 ( K 1 ∗ R 1 ) ( 0 ) = 1 and
∀ t > 0 , ( K 1 ∗ R 1 ) ′ ( t ) = K 1 ′ ( t ) K 1 ( 0 ) + K 1 ( 0 ) ℓ r e s ( t ) + ( K 1 ′ ∗ ℓ r e s ) ( t ) = 0 , \forall t>0,\qquad(K_1 * \mathcal{R}_1)'(t)=\frac{K_1'(t)}{K_1(0)}+K_1(0)\ell_{\mathrm{res}}(t)+(K_1' * \ell_{\mathrm{res}})(t)=0, ∀ t > 0 , ( K 1 ∗ R 1 ) ′ ( t ) = K 1 ( 0 ) K 1 ′ ( t ) + K 1 ( 0 ) ℓ res ( t ) + ( K 1 ′ ∗ ℓ res ) ( t ) = 0 ,
we have K 1 ∗ R 1 = 1 K_1 * \mathcal{R}_1=1 K 1 ∗ R 1 = 1 . The function r m e m r_{\mathrm{mem}} r mem defined in (5.1) therefore satisfies
r m e m = K 1 ′ ∗ R 1 = K 1 ′ K 1 ( 0 ) + K 1 ′ ∗ ℓ r e s = − K 1 ( 0 ) ℓ r e s ∈ C ( R + ) ∩ B V l o c ( R + ) . r_{\mathrm{mem}}=K_1' * \mathcal{R}_1=\frac{K_1'}{K_1(0)}+K_1' * \ell_{\mathrm{res}}=-K_1(0)\ell_{\mathrm{res}}\in C(\mathbb{R}_+) \cap BV_{\mathrm{loc}}(\mathbb{R}_+). r mem = K 1 ′ ∗ R 1 = K 1 ( 0 ) K 1 ′ + K 1 ′ ∗ ℓ res = − K 1 ( 0 ) ℓ res ∈ C ( R + ) ∩ B V loc ( R + ) .
Proof. For Lemma 5.1(i), we fix R ∈ N ∗ R\in\mathbb{N}^* R ∈ N ∗ and set
∀ t ∈ [ 0 , + ∞ ) , U t ( R ) : = s ∫ 0 t 1 { r ≤ τ X ~ R } Γ 1 ( r , X ~ r ) d W r + s ρ ∫ 0 t 1 { r ≤ τ X ~ R } 1 { r ≥ θ } Γ 1 ( r , X ~ r ) σ ( r , X ~ r ) d r . \forall t\in[0,+\infty),\qquad U_t^{(R)}:=s\int_0^t\mathbf{1}_{\{r\leq\tau_{\widetilde X_R}\}}\Gamma_1(r,\widetilde X_r)\,dW_r+s\rho\int_0^t\mathbf{1}_{\{r\leq\tau_{\widetilde X_R}\}}\mathbf{1}_{\{r\geq\theta\}}\Gamma_1(r,\widetilde X_r)\sigma(r,\widetilde X_r)\,dr. ∀ t ∈ [ 0 , + ∞ ) , U t ( R ) := s ∫ 0 t 1 { r ≤ τ X R } Γ 1 ( r , X r ) d W r + s ρ ∫ 0 t 1 { r ≤ τ X R } 1 { r ≥ θ } Γ 1 ( r , X r ) σ ( r , X r ) d r .
For all t ∈ [ 0 , τ ~ R X ] t\in[0,\widetilde\tau_R^X] t ∈ [ 0 , τ R X ] , we have Z ‾ t = ( K 1 ∗ d U ( R ) ) ( t ) \overline Z_t=(K_1*dU^{(R)})(t) Z t = ( K 1 ∗ d U ( R ) ) ( t ) . On each bounded time interval, stopping at τ ~ R X \widetilde\tau_R^X τ R X makes the coefficients bounded. The derivative K 1 ′ K_1' K 1 ′ is also bounded on this interval. Writing
∀ t ≥ 0 , K 1 ( t ) = K 1 ( 0 ) + ∫ 0 t K 1 ′ ( u ) d u , \forall t\geq0,\qquad K_1(t)=K_1(0)+\int_0^tK_1'(u)\,du, ∀ t ≥ 0 , K 1 ( t ) = K 1 ( 0 ) + ∫ 0 t K 1 ′ ( u ) d u ,
the stochastic Fubini theorem [35 Theorem 2.2] gives
∀ t ∈ [ 0 , τ ~ R X ] , Z ‾ t = K 1 ( 0 ) U t ( R ) + ∫ 0 t ( K 1 ′ ∗ d U ( R ) ) ( r ) d r . (67) \forall t\in[0,\widetilde\tau_R^X],\qquad\overline Z_t=K_1(0)U_t^{(R)}+\int_0^t(K_1'*dU^{(R)})(r)\,dr. \tag*{(67)} ∀ t ∈ [ 0 , τ R X ] , Z t = K 1 ( 0 ) U t ( R ) + ∫ 0 t ( K 1 ′ ∗ d U ( R ) ) ( r ) d r . (67)
Since R 1 ∗ K 1 = 1 \mathcal{R}_1*K_1=1 R 1 ∗ K 1 = 1 , we have r m e m ∗ K 1 = K 1 ′ ∗ ( R 1 ∗ K 1 ) = K 1 − K 1 ( 0 ) r_{\mathrm{mem}}*K_1=K_1'*(\mathcal{R}_1*K_1)=K_1-K_1(0) r mem ∗ K 1 = K 1 ′ ∗ ( R 1 ∗ K 1 ) = K 1 − K 1 ( 0 ) . Differentiating gives
K 1 ′ = r m e m ( 0 ) K 1 + d r m e m ∗ K 1 . K_1'=r_{\mathrm{mem}}(0)K_1+dr_{\mathrm{mem}}*K_1. K 1 ′ = r mem ( 0 ) K 1 + d r mem ∗ K 1 .
Applying [2 Lemma 2.1] with kernel K 1 K_1 K 1 and measure d r m e m dr_{\mathrm{mem}} d r mem to the martingale part M ( R ) M^{(R)} M ( R ) of U ( R ) U^{(R)} U ( R ) , whose quadratic variation has a locally bounded density, gives
∀ t ≥ 0 , ( d r m e m ∗ ( K 1 ∗ d M ( R ) ) ) ( t ) = ( ( d r m e m ∗ K 1 ) ∗ d M ( R ) ) ( t ) . \forall t\geq0,\qquad(dr_{\mathrm{mem}}*(K_1*dM^{(R)}))(t)=((dr_{\mathrm{mem}}*K_1)*dM^{(R)})(t). ∀ t ≥ 0 , ( d r mem ∗ ( K 1 ∗ d M ( R ) )) ( t ) = (( d r mem ∗ K 1 ) ∗ d M ( R ) ) ( t ) .
Together with ordinary Fubini for the drift part, this gives up to indistinguishability, for t ∈ [ 0 , τ ~ R X ] t\in[0,\widetilde\tau_R^X] t ∈ [ 0 , τ R X ] ,
( K 1 ′ ∗ d U ( R ) ) ( t ) = r m e m ( 0 ) ( K 1 ∗ d U ( R ) ) ( t ) + ( ( d r m e m ∗ K 1 ) ∗ d U ( R ) ) ( t ) = r m e m ( 0 ) ( K 1 ∗ d U ( R ) ) ( t ) + ( d r m e m ∗ ( K 1 ∗ d U ( R ) ) ) ( t ) = r m e m ( 0 ) Z ‾ t + ( d r m e m ∗ Z ‾ ) ( t ) , \begin{aligned}
(K_1'*dU^{(R)})(t)=r_{\mathrm{mem}}(0)(K_1*dU^{(R)})(t)+((dr_{\mathrm{mem}}*K_1)*dU^{(R)})(t) \\
&=r_{\mathrm{mem}}(0)(K_1*dU^{(R)})(t)+(dr_{\mathrm{mem}}*(K_1*dU^{(R)}))(t) \\
&=r_{\mathrm{mem}}(0)\overline Z_t+(dr_{\mathrm{mem}}*\overline Z)(t),
\end{aligned} ( K 1 ′ ∗ d U ( R ) ) ( t ) = r mem ( 0 ) ( K 1 ∗ d U ( R ) ) ( t ) + (( d r mem ∗ K 1 ) ∗ d U ( R ) ) ( t ) = r mem ( 0 ) ( K 1 ∗ d U ( R ) ) ( t ) + ( d r mem ∗ ( K 1 ∗ d U ( R ) )) ( t ) = r mem ( 0 ) Z t + ( d r mem ∗ Z ) ( t ) ,
since K 1 ∗ d U ( R ) = Z ‾ K_1*dU^{(R)}=\overline Z K 1 ∗ d U ( R ) = Z on [ 0 , τ ~ R X ] [0,\widetilde\tau_R^X] [ 0 , τ R X ] . Together with (B.3), this proves Lemma 5.1(i). The bound in Lemma 5.1(ii) follows from
∀ t ∈ [ 0 , T ‾ ∧ τ ~ R X ] , ∣ ( d r m e m ∗ Z ‾ ) ( t ) ∣ ≤ ∫ ( 0 , t ] ∣ Z ‾ t − r ∣ ∣ d r m e m ∣ ( d r ) ≤ ∣ d r m e m ∣ ( [ 0 , T ‾ ] ) sup u ∈ [ 0 , t ] ∣ Z ‾ u ∣ . \forall t\in[0,\overline T\wedge\widetilde\tau_R^X],\qquad|(dr_{\mathrm{mem}}*\overline Z)(t)|\leq\int_{(0,t]}|\overline Z_{t-r}|\,|dr_{\mathrm{mem}}|(dr)\leq|dr_{\mathrm{mem}}|([0,\overline T])\sup_{u\in[0,t]}|\overline Z_u|. ∀ t ∈ [ 0 , T ∧ τ R X ] , ∣ ( d r mem ∗ Z ) ( t ) ∣ ≤ ∫ ( 0 , t ] ∣ Z t − r ∣ ∣ d r mem ∣ ( d r ) ≤ ∣ d r mem ∣ ([ 0 , T ]) u ∈ [ 0 , t ] sup ∣ Z u ∣.
Appendix C. Proof of the accessibility lemma
We complete the proof of Lemma 2.7 for the polynomial PDV model, specification (A), and the Quadratic Volterra Heston+ model.
Polynomial PDV and specification (A)
We use the assumptions of Theorem 2. For the polynomial PDV model and the model with specification (A), respectively, we have
X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) , σ = X or X , Y 2 = G 2 + K 2 ∗ σ 2 ≥ G 2 ≥ 0. X=\beta+F_1(Y_1)+F_2(Y_2),\qquad\sigma=\sqrt{X}\ \text{or }X,\qquad Y_2=G_2+K_2 * \sigma^2\geq G_2\geq0. X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) , σ = X or X , Y 2 = G 2 + K 2 ∗ σ 2 ≥ G 2 ≥ 0.
Here β \beta β is constant in both models, and we have z 0 = − β 1 2 β 2 ( 1 ) z_0=-\frac{\beta_1}{2\beta_2^{(1)}} z 0 = − 2 β 2 ( 1 ) β 1 or β 0 , 1 − β 0 , 2 β 1 , 2 − β 1 , 1 \frac{\beta_{0,1}-\beta_{0,2}}{\beta_{1,2}-\beta_{1,1}} β 1 , 2 − β 1 , 1 β 0 , 1 − β 0 , 2 . In both models, we have Γ 1 = σ \Gamma_1=\sigma Γ 1 = σ and Γ 2 = σ 2 \Gamma_2=\sigma^2 Γ 2 = σ 2 . The function F 1 F_1 F 1 decreases on ( − ∞ , z 0 ] (-\infty,z_0] ( − ∞ , z 0 ] and has its global minimum at z 0 z_0 z 0 . Also, we have lim y → − ∞ F 1 ( y ) = + ∞ \lim_{y\to-\infty}F_1(y)=+\infty lim y → − ∞ F 1 ( y ) = + ∞ , and F 2 F_2 F 2 is nondecreasing and unbounded above. We assume that X 0 > 0 X_0>0 X 0 > 0 and fix T > 0 T>0 T > 0 . If Y 1 , 0 > z 0 Y_{1,0}>z_0 Y 1 , 0 > z 0 , then X 0 > 0 X_0>0 X 0 > 0 gives τ ∗ = 0 < T \tau_*=0<T τ ∗ = 0 < T . We now consider Y 1 , 0 ≤ z 0 Y_{1,0}\leq z_0 Y 1 , 0 ≤ z 0 .
Remark C.1. Moving directly towards z 0 z_0 z 0 may bring X X X to zero. We therefore choose a deterministic reference path for Y 1 Y_1 Y 1 that first takes sufficiently negative values and then reaches a value above z 0 z_0 z 0 . Since X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) X=\beta+F_1(Y_1)+F_2(Y_2) X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) , large negative values of Y 1 Y_1 Y 1 make β + F 1 ( Y 1 ) \beta+F_1(Y_1) β + F 1 ( Y 1 ) , X X X , and σ 2 \sigma^2 σ 2 large. This creates a large contribution to Y 2 Y_2 Y 2 through K 2 ∗ σ 2 K_2 * \sigma^2 K 2 ∗ σ 2 . If the reference path moves from these negative values to a value above z 0 z_0 z 0 in a sufficiently short time, as quantified in (C.3), this contribution keeps F 2 ( Y 2 ) F_2(Y_2) F 2 ( Y 2 ) large enough so that X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) ≥ β + F 1 ( z 0 ) + F 2 ( Y 2 ) > 0 X=\beta+F_1(Y_1)+F_2(Y_2)\geq\beta+F_1(z_0)+F_2(Y_2)>0 X = β + F 1 ( Y 1 ) + F 2 ( Y 2 ) ≥ β + F 1 ( z 0 ) + F 2 ( Y 2 ) > 0 . Thus the path crosses z 0 z_0 z 0 while the diffusion coefficient stays nonzero, with Γ 1 = σ ≠ 0 \Gamma_1=\sigma\ne0 Γ 1 = σ = 0 .
For the polynomial PDV model, the quadratic expression for X X X in (11) gives
∀ t ∈ [ 0 , T ] , σ t 2 = X t ≤ C ( 1 + Y 1 , t 2 ) + C ∣ Y 2 , t ∣ . \forall t\in[0,T],\qquad\sigma_t^2=X_t\leq C(1+Y_{1,t}^2)+C|Y_{2,t}|. ∀ t ∈ [ 0 , T ] , σ t 2 = X t ≤ C ( 1 + Y 1 , t 2 ) + C ∣ Y 2 , t ∣.
For specification (A), the linear growth of F 1 F_1 F 1 gives
∀ t ∈ [ 0 , T ] , σ t 2 = ( β + F 1 ( Y 1 , t ) + β 2 Y 2 , t ) 2 \forall t\in[0,T],\qquad\sigma_t^2=\left(\beta+F_1(Y_{1,t})+\beta_2\sqrt{Y_{2,t}}\right)^2 ∀ t ∈ [ 0 , T ] , σ t 2 = ( β + F 1 ( Y 1 , t ) + β 2 Y 2 , t ) 2
≤ 2 ( β + F 1 ( Y 1 , t ) ) 2 + 2 β 2 2 Y 2 , t ≤ C ( 1 + Y 1 , t 2 ) + C ∣ Y 2 , t ∣ . \leq2\left(\beta+F_1(Y_{1,t})\right)^2+2\beta_2^2Y_{2,t}\leq C(1+Y_{1,t}^2)+C|Y_{2,t}|. ≤ 2 ( β + F 1 ( Y 1 , t ) ) 2 + 2 β 2 2 Y 2 , t ≤ C ( 1 + Y 1 , t 2 ) + C ∣ Y 2 , t ∣.
In both models, Y 2 = G 2 + K 2 ∗ σ 2 Y_2=G_2+K_2 * \sigma^2 Y 2 = G 2 + K 2 ∗ σ 2 and K 2 ≥ 0 K_2\geq0 K 2 ≥ 0 , so we have
∀ t ∈ [ 0 , T ] , ∣ Y 2 , t ∣ ≤ ∥ G 2 ∥ ∞ , [ 0 , T ] + ∫ 0 t K 2 ( t − s ) σ s 2 d s . \forall t\in[0,T],\qquad|Y_{2,t}|\leq\|G_2\|_{\infty,[0,T]}+\int_0^t K_2(t-s)\sigma_s^2\,\mathrm{d}s. ∀ t ∈ [ 0 , T ] , ∣ Y 2 , t ∣ ≤ ∥ G 2 ∥ ∞ , [ 0 , T ] + ∫ 0 t K 2 ( t − s ) σ s 2 d s .
Substituting the bound on ∣ Y 2 , t ∣ |Y_{2,t}| ∣ Y 2 , t ∣ into the corresponding bound on σ t 2 \sigma_t^2 σ t 2 gives, for some C > 0 C>0 C > 0 ,
∀ t ∈ [ 0 , T ] , σ t 2 ≤ C ( 1 + Y 1 , t 2 ) + C ∫ 0 t K 2 ( t − s ) σ s 2 d s . (68) \forall t\in[0,T],\qquad\sigma_t^2\leq C(1+Y_{1,t}^2)+C\int_0^t K_2(t-s)\sigma_s^2\,\mathrm{d}s. \tag*{(68)} ∀ t ∈ [ 0 , T ] , σ t 2 ≤ C ( 1 + Y 1 , t 2 ) + C ∫ 0 t K 2 ( t − s ) σ s 2 d s . (68)
Step 1: Moving to the left with positive volatility. The map ( s , a ) ↦ β + F 1 ( ( G 1 ( s ) + a ) ∧ z 0 ) + F 2 ( G 2 ( s ) ) (s,a)\mapsto\beta+F_1((G_1(s)+a)\wedge z_0)+F_2(G_2(s)) ( s , a ) ↦ β + F 1 (( G 1 ( s ) + a ) ∧ z 0 ) + F 2 ( G 2 ( s )) is continuous at ( 0 , 0 ) (0,0) ( 0 , 0 ) , where its value is X 0 > 0 X_0>0 X 0 > 0 . Hence there exist b ∈ ( 0 , T 2 ) b\in(0,\frac{T}{2}) b ∈ ( 0 , 2 T ) and δ ∈ ( 0 , 1 2 ] \delta\in(0,\frac{1}{2}] δ ∈ ( 0 , 2 1 ] such that
∀ s ∈ [ 0 , b ] , β + F 1 ( ( G 1 ( s ) + δ ) ∧ z 0 ) + F 2 ( G 2 ( s ) ) ≥ X 0 2 . (69) \forall s\in[0,b],\qquad\beta+F_1((G_1(s)+\delta)\wedge z_0)+F_2(G_2(s))\geq\frac{X_0}{2}. \tag*{(69)} ∀ s ∈ [ 0 , b ] , β + F 1 (( G 1 ( s ) + δ ) ∧ z 0 ) + F 2 ( G 2 ( s )) ≥ 2 X 0 . (69)
Continuity of v ↦ ∫ v v + b 2 K 2 ( u ) d u v\mapsto\int_v^{v+\frac{b}{2}}K_2(u)\,\mathrm{d}u v ↦ ∫ v v + 2 b K 2 ( u ) d u and the inequality ∫ 0 b 2 K 2 ( u ) d u > 0 \int_0^{\frac{b}{2}}K_2(u)\,\mathrm{d}u>0 ∫ 0 2 b K 2 ( u ) d u > 0 from Assumption 1 give h ∈ ( 0 , T − b ) h\in(0,T-b) h ∈ ( 0 , T − b ) such that
inf v ∈ [ 0 , h ] ∫ v v + b 2 K 2 ( u ) d u ≥ 1 2 ∫ 0 b 2 K 2 ( u ) d u > 0. (70) \inf_{v\in[0,h]}\int_v^{v+\frac{b}{2}}K_2(u)\,\mathrm{d}u\geq\frac{1}{2}\int_0^{\frac{b}{2}}K_2(u)\,\mathrm{d}u>0. \tag*{(70)} v ∈ [ 0 , h ] inf ∫ v v + 2 b K 2 ( u ) d u ≥ 2 1 ∫ 0 2 b K 2 ( u ) d u > 0. (70)
For R ≥ 0 R \ge0 R ≥ 0 , the function
∀ s ∈ [ 0 , T ] , p ( s ) = − R s + ( ( s − b ) + ) 2 h 2 ( z 0 + 1 − G 1 ( b + h ) + R ( b + h ) ) \forall s \in[0,T],\qquad p(s)=-Rs+\frac{((s-b)^+)^2}{h^2}\left(z_0+1-G_1(b+h)+R(b+h)\right) ∀ s ∈ [ 0 , T ] , p ( s ) = − R s + h 2 (( s − b ) + ) 2 ( z 0 + 1 − G 1 ( b + h ) + R ( b + h ) )
belongs to C 1 ( [ 0 , T ] ) C^1([0,T]) C 1 ([ 0 , T ]) , equals − R s -Rs − R s on [ 0 , b ] [0,b] [ 0 , b ] , and satisfies G 1 ( b + h ) + p ( b + h ) = z 0 + 1 G_1(b+h)+p(b+h)=z_0+1 G 1 ( b + h ) + p ( b + h ) = z 0 + 1 . We define
τ : = ( b + h ) ∧ inf { s ∈ [ 0 , b + h ] : ∣ Y 1 , s − G 1 ( s ) − p ( s ) ∣ ≥ δ } . \tau:=(b+h)\wedge\inf\{s\in[0,b+h]:\lvert Y_{1,s}-G_1(s)-p(s)\rvert\ge\delta\}. τ := ( b + h ) ∧ inf { s ∈ [ 0 , b + h ] : ∣ Y 1 , s − G 1 ( s ) − p ( s )∣ ≥ δ } .
Since p ( 0 ) = 0 p(0)=0 p ( 0 ) = 0 and Y 1 − G 1 − p Y_1-G_1-p Y 1 − G 1 − p is continuous, we have
sup s ∈ [ 0 , τ ] ∣ Y 1 , s − G 1 ( s ) − p ( s ) ∣ ≤ δ . \sup_{s\in[0,\tau]}\lvert Y_{1,s}-G_1(s)-p(s)\rvert\le\delta. s ∈ [ 0 , τ ] sup ∣ Y 1 , s − G 1 ( s ) − p ( s )∣ ≤ δ .
In particular, we have
∀ s ∈ [ 0 , b ∧ τ ] , Y 1 , s ≤ G 1 ( s ) − R s + δ ≤ G 1 ( s ) + δ . (71) \forall s\in[0,b\wedge\tau],\qquad Y_{1,s}\le G_1(s)-Rs+\delta\le G_1(s)+\delta. \tag*{(71)} ∀ s ∈ [ 0 , b ∧ τ ] , Y 1 , s ≤ G 1 ( s ) − R s + δ ≤ G 1 ( s ) + δ . (71)
Since F 1 F_1 F 1 is nonincreasing on ( − ∞ , z 0 ] (-\infty,z_0] ( − ∞ , z 0 ] and has its global minimum at z 0 z_0 z 0 , we have
∀ s ∈ [ 0 , b ∧ τ ] , F 1 ( Y 1 , s ) ≥ F 1 ( ( G 1 ( s ) + δ ) ∧ z 0 ) . \forall s\in[0,b\wedge\tau],\qquad F_1(Y_{1,s})\ge F_1\left((G_1(s)+\delta)\wedge z_0\right). ∀ s ∈ [ 0 , b ∧ τ ] , F 1 ( Y 1 , s ) ≥ F 1 ( ( G 1 ( s ) + δ ) ∧ z 0 ) .
Also, Y 2 ≥ G 2 Y_2\ge G_2 Y 2 ≥ G 2 and F 2 F_2 F 2 is nondecreasing. Hence, by the initial lower bound (C.2), we have
∀ s ∈ [ 0 , b ∧ τ ] , X s ≥ β + F 1 ( ( G 1 ( s ) + δ ) ∧ z 0 ) + F 2 ( G 2 ( s ) ) ≥ X 0 2 . \forall s\in[0,b\wedge\tau],\qquad X_s\ge\beta+F_1\left((G_1(s)+\delta)\wedge z_0\right)+F_2(G_2(s))\ge\frac{X_0}{2}. ∀ s ∈ [ 0 , b ∧ τ ] , X s ≥ β + F 1 ( ( G 1 ( s ) + δ ) ∧ z 0 ) + F 2 ( G 2 ( s )) ≥ 2 X 0 .
On { τ < b } \{\tau<b\} { τ < b } , we have X s ≥ X 0 2 X_s\ge\frac{X_0}{2} X s ≥ 2 X 0 for every s ∈ [ 0 , τ ] s\in[0,\tau] s ∈ [ 0 , τ ] . On { τ ≥ b } \{\tau\ge b\} { τ ≥ b } , for all sufficiently large R R R , the bound on Y 1 Y_1 Y 1 in (C.4) gives
sup s ∈ [ b 2 , b ] Y 1 , s ≤ ∥ G 1 ∥ ∞ , [ 0 , b ] − R b 2 + δ < z 0 , \sup_{s\in[\frac b2,b]}Y_{1,s}\le\lVert G_1\rVert_{\infty,[0,b]}-\frac{Rb}{2}+\delta<z_0, s ∈ [ 2 b , b ] sup Y 1 , s ≤ ∥ G 1 ∥ ∞ , [ 0 , b ] − 2 R b + δ < z 0 ,
inf s ∈ [ b 2 , b ] X s ≥ β + F 1 ( ∥ G 1 ∥ ∞ , [ 0 , b ] − R b 2 + δ ) + min s ∈ [ 0 , b ] F 2 ( G 2 ( s ) ) → R → ∞ + ∞ . \inf_{s\in[\frac b2,b]}X_s\ge\beta+F_1\left(\lVert G_1\rVert_{\infty,[0,b]}-\frac{Rb}{2}+\delta\right)+\min_{s\in[0,b]}F_2(G_2(s))\xrightarrow[R\to\infty]{}+\infty. s ∈ [ 2 b , b ] inf X s ≥ β + F 1 ( ∥ G 1 ∥ ∞ , [ 0 , b ] − 2 R b + δ ) + s ∈ [ 0 , b ] min F 2 ( G 2 ( s )) R → ∞ + ∞.
For sufficiently large R R R , this lower bound is at least 1 1 1 . Thus we have σ s 2 ≥ X s \sigma_s^2\ge X_s σ s 2 ≥ X s on [ b 2 , b ] [\frac b2,b] [ 2 b , b ] in both models, and we have
inf s ∈ [ b 2 , b ] σ s 2 ≥ β + F 1 ( ∥ G 1 ∥ ∞ , [ 0 , b ] − R b 2 + δ ) + min s ∈ [ 0 , b ] F 2 ( G 2 ( s ) ) → R → ∞ + ∞ . (72) \inf_{s\in[\frac b2,b]}\sigma_s^2\ge\beta+F_1\left(\lVert G_1\rVert_{\infty,[0,b]}-\frac{Rb}{2}+\delta\right)+\min_{s\in[0,b]}F_2(G_2(s))\xrightarrow[R\to\infty]{}+\infty. \tag*{(72)} s ∈ [ 2 b , b ] inf σ s 2 ≥ β + F 1 ( ∥ G 1 ∥ ∞ , [ 0 , b ] − 2 R b + δ ) + s ∈ [ 0 , b ] min F 2 ( G 2 ( s )) R → ∞ + ∞. (72)
Step 2: Crossing z 0 z_0 z 0 while keeping X X X positive. On { τ ≥ b } \{\tau\ge b\} { τ ≥ b } , for every v ∈ [ 0 , h ] v\in[0,h] v ∈ [ 0 , h ] , the lower bound on the integral of K 2 K_2 K 2 in (C.3) gives
∫ b 2 b K 2 ( b + v − s ) d s = ∫ v v + b 2 K 2 ( u ) d u ≥ 1 2 ∫ 0 b 2 K 2 ( u ) d u . \int_{\frac b2}^{b}K_2(b+v-s)\,\mathrm{d}s=\int_v^{v+\frac b2}K_2(u)\,\mathrm{d}u\ge\frac12\int_0^{\frac b2}K_2(u)\,\mathrm{d}u. ∫ 2 b b K 2 ( b + v − s ) d s = ∫ v v + 2 b K 2 ( u ) d u ≥ 2 1 ∫ 0 2 b K 2 ( u ) d u .
Consequently, we have
Y 2 , b + v ≥ G 2 ( b + v ) + ∫ b 2 b K 2 ( b + v − s ) σ s 2 d s ≥ min s ∈ [ 0 , T ] G 2 ( s ) + 1 2 ( inf s ∈ [ b 2 , b ] σ s 2 ) ∫ 0 b 2 K 2 ( u ) d u . Y_{2,b+v}\ge G_2(b+v)+\int_{\frac b2}^{b}K_2(b+v-s)\sigma_s^2\,\mathrm{d}s\ge\min_{s\in[0,T]}G_2(s)+\frac12\left(\inf_{s\in[\frac b2,b]}\sigma_s^2\right)\int_0^{\frac b2}K_2(u)\,\mathrm{d}u. Y 2 , b + v ≥ G 2 ( b + v ) + ∫ 2 b b K 2 ( b + v − s ) σ s 2 d s ≥ s ∈ [ 0 , T ] min G 2 ( s ) + 2 1 ( s ∈ [ 2 b , b ] inf σ s 2 ) ∫ 0 2 b K 2 ( u ) d u .
By (C.5), since F 1 ≥ F 1 ( z 0 ) F_1\ge F_1(z_0) F 1 ≥ F 1 ( z 0 ) and F 2 ( y ) → y → + ∞ + ∞ F_2(y)\xrightarrow[y\to+\infty]{}+\infty F 2 ( y ) y → + ∞ + ∞ , there exists R R R large enough that, on { τ ≥ b } \{\tau\ge b\} { τ ≥ b } ,
∀ v ∈ [ 0 , h ] , X b + v ≥ β + F 1 ( z 0 ) + F 2 ( Y 2 , b + v ) ≥ 1. (73) \forall v\in[0,h],\qquad X_{b+v}\ge\beta+F_1(z_0)+F_2(Y_{2,b+v})\ge1. \tag*{(73)} ∀ v ∈ [ 0 , h ] , X b + v ≥ β + F 1 ( z 0 ) + F 2 ( Y 2 , b + v ) ≥ 1. (73)
We keep this choice of R R R , p p p , and τ \tau τ . Then we have
∀ s ∈ [ 0 , τ ] , X s ≥ X 0 2 ∧ 1. \forall s\in[0,\tau],\qquad X_s\ge\frac{X_0}{2}\wedge1. ∀ s ∈ [ 0 , τ ] , X s ≥ 2 X 0 ∧ 1.
Step 3: Applying the tube estimate. The definition of τ \tau τ gives
sup s ∈ [ 0 , τ ] ∣ Y 1 , s ∣ ≤ ∥ G 1 ∥ ∞ , [ 0 , T ] + ∥ p ∥ ∞ , [ 0 , T ] + δ . \sup_{s\in[0,\tau]} |Y_{1,s}| \leq\|G_1\|_{\infty,[0,T]}+\|p\|_{\infty,[0,T]}+\delta. s ∈ [ 0 , τ ] sup ∣ Y 1 , s ∣ ≤ ∥ G 1 ∥ ∞ , [ 0 , T ] + ∥ p ∥ ∞ , [ 0 , T ] + δ .
The Volterra–Grönwall lemma [37 Lemma 2.2] applied to the bound on σ 2 \sigma^2 σ 2 in (C.1) on [ 0 , τ ] [0,\tau] [ 0 , τ ] gives a finite constant C C C , depending on p p p , δ \delta δ , T T T and the initial curves, such that
sup s ∈ [ 0 , τ ] ∣ σ s ∣ ≤ C . (74) \sup_{s\in[0,\tau]} |\sigma_s| \leq C. \tag*{(74)} s ∈ [ 0 , τ ] sup ∣ σ s ∣ ≤ C . (74)
Therefore, for some 0 < c ≤ C < ∞ 0<c\leq C<\infty 0 < c ≤ C < ∞ , we have
∀ s ∈ [ 0 , τ ] , c ≤ ∣ Γ 1 ( s , X s ) ∣ = ∣ σ s ∣ ≤ C . \forall s\in[0,\tau], \qquad c\leq|\Gamma_1(s,X_s)|=|\sigma_s|\leq C. ∀ s ∈ [ 0 , τ ] , c ≤ ∣ Γ 1 ( s , X s ) ∣ = ∣ σ s ∣ ≤ C .
As in the proof of Lemma 5.1, the model (PDV) gives the semimartingale representation
∀ t ∈ [ 0 , τ ] , Y 1 , t − G 1 ( t ) = K 1 ( 0 ) ∫ 0 t Γ 1 ( u , X u ) d W u + ∫ 0 t [ r m e m ( 0 ) ( Y 1 , u − G 1 ( u ) ) + ( d r m e m ∗ ( Y 1 − G 1 ) ) ( u ) ] d u . (75) \forall t\in[0,\tau], \qquad Y_{1,t}-G_1(t)=K_1(0)\int_0^t\Gamma_1(u,X_u)\,\mathrm{d}W_u
+\int_0^t\left[r_{\mathrm{mem}}(0)(Y_{1,u}-G_1(u))+(dr_{\mathrm{mem}}*(Y_1-G_1))(u)\right]\mathrm{d}u. \tag*{(75)} ∀ t ∈ [ 0 , τ ] , Y 1 , t − G 1 ( t ) = K 1 ( 0 ) ∫ 0 t Γ 1 ( u , X u ) d W u + ∫ 0 t [ r mem ( 0 ) ( Y 1 , u − G 1 ( u )) + ( d r mem ∗ ( Y 1 − G 1 )) ( u ) ] d u . (75)
The bound (46) , with Y 1 − G 1 Y_1-G_1 Y 1 − G 1 in place of Z ˉ \bar{Z} Z ˉ , bounds its drift on [ 0 , τ ] [0,\tau] [ 0 , τ ] . Lemma 5.2, applied to Y 1 − G 1 Y_1-G_1 Y 1 − G 1 on [ 0 , b + h ] [0,b+h] [ 0 , b + h ] , with stopping time τ \tau τ , gives
P 0 ( sup s ∈ [ 0 , τ ] ∣ Y 1 , s − G 1 ( s ) − p ( s ) ∣ < δ ) > 0. \mathbb{P}_0\left(\sup_{s\in[0,\tau]}|Y_{1,s}-G_1(s)-p(s)|<\delta\right)>0. P 0 ( s ∈ [ 0 , τ ] sup ∣ Y 1 , s − G 1 ( s ) − p ( s ) ∣ < δ ) > 0.
On { τ < b + h } \{\tau<b+h\} { τ < b + h } , continuity of Y 1 − G 1 − p Y_1-G_1-p Y 1 − G 1 − p gives ∣ Y 1 , τ − G 1 ( τ ) − p ( τ ) ∣ = δ |Y_{1,\tau}-G_1(\tau)-p(\tau)|=\delta ∣ Y 1 , τ − G 1 ( τ ) − p ( τ ) ∣ = δ . Therefore we have P 0 ( τ = b + h ) > 0 \mathbb{P}_0(\tau=b+h)>0 P 0 ( τ = b + h ) > 0 . On { τ = b + h } \{\tau=b+h\} { τ = b + h } , (C.6) gives X b + h ≥ 1 X_{b+h}\geq1 X b + h ≥ 1 , and the definition of τ \tau τ gives
Y 1 , b + h ≥ G 1 ( b + h ) + p ( b + h ) − δ = z 0 + 1 − δ ≥ z 0 + 1 2 . Y_{1,b+h}\geq G_1(b+h)+p(b+h)-\delta=z_0+1-\delta\geq z_0+\frac{1}{2}. Y 1 , b + h ≥ G 1 ( b + h ) + p ( b + h ) − δ = z 0 + 1 − δ ≥ z 0 + 2 1 .
Hence we have
P 0 ( X b + h > 0 , Y 1 , b + h > z 0 ) ≥ P 0 ( τ = b + h ) > 0. (76) \mathbb{P}_0(X_{b+h}>0,Y_{1,b+h}>z_0)\geq\mathbb{P}_0(\tau=b+h)>0. \tag*{(76)} P 0 ( X b + h > 0 , Y 1 , b + h > z 0 ) ≥ P 0 ( τ = b + h ) > 0. (76)
Since s = ε s = 1 s=\varepsilon_s=1 s = ε s = 1 , x ∗ = 0 x^*=0 x ∗ = 0 , and b + h < T b+h<T b + h < T , (C.9) gives
P 0 ( τ ∗ < T ) ≥ P 0 ( X b + h > 0 , Y 1 , b + h > z 0 ) > 0 , \mathbb{P}_0(\tau^*<T)\geq\mathbb{P}_0(X_{b+h}>0,Y_{1,b+h}>z_0)>0, P 0 ( τ ∗ < T ) ≥ P 0 ( X b + h > 0 , Y 1 , b + h > z 0 ) > 0 ,
which proves (25) for these two models.
Quadratic Volterra Heston+
For the Quadratic Volterra Heston+ model with a 0 > 0 a_0>0 a 0 > 0 and 1 ≤ p ≤ 2 1\leq p\leq2 1 ≤ p ≤ 2 , under the assumptions of Corollary 2.6(ii), we have X = Y 1 X=Y_1 X = Y 1 , so X − G 1 X-G_1 X − G 1 has the semimartingale representation (C.8). For t > 0 t>0 t > 0 , a ∈ R a\in\mathbb{R} a ∈ R , and δ > 0 \delta>0 δ > 0 , we define the path p p p by
∀ s ∈ [ 0 , t ] , p ( s ) : = s t ( a − G 1 ( t ) ) , \forall s\in[0,t], \qquad p(s):=\frac{s}{t}(a-G_1(t)), ∀ s ∈ [ 0 , t ] , p ( s ) := t s ( a − G 1 ( t )) ,
and consider the tube
{ ( u , x ) ∈ [ 0 , t ] × R : ∣ x − G 1 ( u ) − p ( u ) ∣ < δ } . (77) \{(u,x)\in[0,t]\times\mathbb{R}:|x-G_1(u)-p(u)|<\delta\}. \tag*{(77)} {( u , x ) ∈ [ 0 , t ] × R : ∣ x − G 1 ( u ) − p ( u ) ∣ < δ } . (77)
We set
τ : = t ∧ inf { u ∈ [ 0 , t ] : ∣ X u − G 1 ( u ) − p ( u ) ∣ ≥ δ } . \tau:=t\wedge\inf\{u\in[0,t]:|X_u-G_1(u)-p(u)|\geq\delta\}. τ := t ∧ inf { u ∈ [ 0 , t ] : ∣ X u − G 1 ( u ) − p ( u ) ∣ ≥ δ } .
On [ 0 , τ ] [0,\tau] [ 0 , τ ] , Γ 1 = V \Gamma_1=\sqrt{V} Γ 1 = V is bounded and bounded below by a 0 > 0 \sqrt{a_0}>0 a 0 > 0 . On [ 0 , τ ] [0,\tau] [ 0 , τ ] , the drift of X − G 1 X-G_1 X − G 1 is also bounded by (5.3) of Lemma 5.1. If sup u ∈ [ 0 , τ ] ∣ X u − G 1 ( u ) − p ( u ) ∣ < δ \sup_{u\in[0,\tau]}|X_u-G_1(u)-p(u)|<\delta sup u ∈ [ 0 , τ ] ∣ X u − G 1 ( u ) − p ( u ) ∣ < δ , continuity gives τ = t \tau=t τ = t , and ∣ X t − a ∣ < δ |X_t-a|<\delta ∣ X t − a ∣ < δ follows from G 1 ( t ) + p ( t ) = a G_1(t)+p(t)=a G 1 ( t ) + p ( t ) = a . Lemma 5.2 , applied to the tube (C.10) up to τ \tau τ , therefore gives
∀ t > 0 , ∀ a ∈ R , ∀ δ > 0 , P 0 ( ∣ X t − a ∣ < δ ) ≥ P 0 ( sup u ∈ [ 0 , τ ] ∣ X u − G 1 ( u ) − p ( u ) ∣ < δ ) > 0. \forall t>0,\quad\forall a\in\mathbb{R},\quad\forall\delta>0,\quad\mathbb{P}_0(|X_t-a|<\delta)\geq\mathbb{P}_0\left(\sup_{u\in[0,\tau]}|X_u-G_1(u)-p(u)|<\delta\right)>0. ∀ t > 0 , ∀ a ∈ R , ∀ δ > 0 , P 0 ( ∣ X t − a ∣ < δ ) ≥ P 0 ( u ∈ [ 0 , τ ] sup ∣ X u − G 1 ( u ) − p ( u ) ∣ < δ ) > 0.
In particular, we have P 0 ( X t > 0 ) > 0 \mathbb{P}_0(X_t>0)>0 P 0 ( X t > 0 ) > 0 and P 0 ( X t < 0 ) > 0 \mathbb{P}_0(X_t<0)>0 P 0 ( X t < 0 ) > 0 for every t > 0 t>0 t > 0 . Here X t = Y 1 , t X_t=Y_{1,t} X t = Y 1 , t and ε s = s \varepsilon_s=s ε s = s , so the full-support estimate also gives
∀ t > 0 , P 0 ( ε s X t > x ∗ , s Y 1 , t > z 0 ) = P 0 ( s X t > max { x ∗ , z 0 } ) > 0. \forall t>0,\qquad\mathbb{P}_0(\varepsilon_sX_t>x^*,\ sY_{1,t}>z_0)=\mathbb{P}_0(sX_t>\max\{x^*,z_0\})>0. ∀ t > 0 , P 0 ( ε s X t > x ∗ , s Y 1 , t > z 0 ) = P 0 ( s X t > max { x ∗ , z 0 }) > 0.
Taking t = T / 2 t=T/2 t = T /2 gives P 0 ( τ ∗ < T ) > 0 \mathbb{P}_0(\tau_*<T)>0 P 0 ( τ ∗ < T ) > 0 , completing the proof of Lemma 2.7 .
References