Introduction
Fix a time horizon T > 0 T>0 T > 0 and an observation filtration F = ( F t ) 0 ≤ t ≤ T \mathbf{F}=(\mathcal{F}_t)_{0\le t\le T} F = ( F t ) 0 ≤ t ≤ T , where F t \mathcal{F}_t F t represents the information available at time t t t . A controller tracks a continuous adapted target Θ = ( Θ t ) 0 ≤ t ≤ T \Theta=(\Theta_t)_{0\le t\le T} Θ = ( Θ t ) 0 ≤ t ≤ T by choosing a control ϕ = ( ϕ t ) 0 ≤ t ≤ T \phi=(\phi_t)_{0\le t\le T} ϕ = ( ϕ t ) 0 ≤ t ≤ T . Adaptedness means that the target and control at time t t t are measurable with respect to F t \mathcal{F}_t F t . The positive deterministic weights a t a_t a t and b t b_t b t specify the relative importance of tracking error and adjustments. The parameter λ > 0 \lambda>0 λ > 0 sets the overall adjustment-cost level. Writing E \mathbb{E} E for expectation, we minimize
J λ ( ϕ ; Θ ) : = E [ 1 2 ∫ 0 T a t ( ϕ t − Θ t ) 2 d t + λ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] . (1) J_\lambda(\phi;\Theta):=\mathbb{E}\left[\frac{1}{2}\int_0^T a_t(\phi_t-\Theta_t)^2\,\mathrm{d}t+\lambda\int_{[0,T]}b_t\,\mathrm{d}|\phi|_t\right]. \tag*{(1)} J λ ( ϕ ; Θ ) := E [ 2 1 ∫ 0 T a t ( ϕ t − Θ t ) 2 d t + λ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] . (1)
We minimize over adapted, càdlàg controls of finite variation and write
V λ ( Θ ) : = inf ϕ J λ ( ϕ ; Θ ) . V_\lambda(\Theta):=\inf_{\phi}J_\lambda(\phi;\Theta). V λ ( Θ ) := ϕ inf J λ ( ϕ ; Θ ) .
Here càdlàg means right-continuous with left limits, and finite variation means that the total amount of adjustment is finite. The notation d ∣ ϕ ∣ t \mathrm{d}|\phi|_t d ∣ ϕ ∣ t denotes the total-variation measure of the control increments. It includes the initial adjustment from the prescribed pre-initial position ϕ 0 − = Θ 0 \phi_{0-}=\Theta_0 ϕ 0 − = Θ 0 , with cost λ b 0 ∣ ϕ 0 − − Θ 0 ∣ \lambda b_0|\phi_{0-}-\Theta_0| λ b 0 ∣ ϕ 0 − − Θ 0 ∣ . § 2 gives the precise admissible class, integrability requirements and coefficient assumptions. Suppose that target increments over a small time interval of length h > 0 h>0 h > 0 have typical size h H h^H h H , where H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) describes their scaling. An error tolerance δ > 0 \delta>0 δ > 0 then suggests an adjustment interval h ≍ δ 1 / H h\asymp\delta^{1/H} h ≍ δ 1/ H . We use ≍ \asymp ≍ for comparison up to positive multiplicative constants independent of the small parameter. The tracking cost per unit time is then of order δ 2 \delta^2 δ 2 , whereas the variation-cost rate satisfies
λ δ h ≍ λ δ 1 − 1 / H . \lambda\frac{\delta}{h}\asymp\lambda\delta^{1-1/H}. λ h δ ≍ λ δ 1 − 1/ H .
Balancing these two rates suggests a tracking-error scale δ λ \delta_\lambda δ λ and a value of order
δ λ ≍ λ H / ( 1 + H ) , V λ ( Θ ) ≍ λ 2 H / ( 1 + H ) . (2) \delta_\lambda\asymp\lambda^{H/(1+H)}, \qquad V_\lambda(\Theta) \asymp\lambda^{2H/(1+H)}. \tag*{(2)} δ λ ≍ λ H / ( 1 + H ) , V λ ( Θ ) ≍ λ 2 H / ( 1 + H ) . (2)
An increment–energy inequality proves the lower bound even for noncausal controls , which have access to future information. Foresight therefore leaves the exponent unchanged. For the fractional Brownian target, let B H B^H B H be standard fractional Brownian motion with Hurst parameter H H H , normalized by B 0 H = 0 B^H_0=0 B 0 H = 0 and
E [ ( B t + h H − B t H ) 2 ] = h 2 H , t ≥ 0 , h > 0. \mathbb{E}\left[(B^H_{t+h}-B^H_t)^2\right]=h^{2H}, \qquad t\geq0,\ h>0. E [ ( B t + h H − B t H ) 2 ] = h 2 H , t ≥ 0 , h > 0.
The constant σ > 0 \sigma>0 σ > 0 scales the target amplitude. For Θ = σ B H \Theta=\sigma B^H Θ = σ B H and constant weights a t = a > 0 a_t=a>0 a t = a > 0 and b t = b > 0 b_t=b>0 b t = b > 0 , we write
V λ ( T ; a , b , σ ) : = V λ ( σ B H ) , V_\lambda(T;a,b,\sigma):=V_\lambda(\sigma B^H), V λ ( T ; a , b , σ ) := V λ ( σ B H ) ,
using the usual natural filtration of B H B^H B H , so that controls use only observations of the target from time zero up to the current time. The natural time scale is
ℓ λ : = ( λ b a σ ) 1 / ( 1 + H ) . (3) \ell_\lambda:=\left(\frac{\lambda b}{a\sigma}\right)^{1/(1+H)}. \tag*{(3)} ℓ λ := ( aσ λb ) 1/ ( 1 + H ) . (3)
Rescaling time, the target, the control and the cost transforms the small-cost problem into a normalized tracking problem, called the cell problem , with a = b = σ = λ = 1 a=b=\sigma=\lambda=1 a = b = σ = λ = 1 . Its horizon is L : = T / ℓ λ L:=T/\ell_\lambda L := T / ℓ λ , which tends to infinity as λ → 0 \lambda\to0 λ → 0 . The limiting optimal cost per unit time is denoted by κ H \kappa_H κ H . A terminal reset is a final adjustment of the control to the target. Adding its cost makes the cell value subadditive, and the extra cost is sublinear in L L L . Consequently,
V λ ( T ; a , b , σ ) ∼ λ → 0 κ H T a ( 1 − H ) / ( 1 + H ) ( λ b ) 2 H / ( 1 + H ) σ 2 / ( 1 + H ) , κ H ∈ ( 0 , ∞ ) . (4) V_\lambda(T;a,b,\sigma)\mathop{\sim}_{\lambda\to0}\kappa_H T a^{(1-H)/(1+H)}(\lambda b)^{2H/(1+H)}\sigma^{2/(1+H)}, \qquad\kappa_H\in(0,\infty). \tag*{(4)} V λ ( T ; a , b , σ ) ∼ λ → 0 κ H T a ( 1 − H ) / ( 1 + H ) ( λb ) 2 H / ( 1 + H ) σ 2/ ( 1 + H ) , κ H ∈ ( 0 , ∞ ) . (4)
Full past information means that the controller also observes the target at all negative times, using a two-sided fractional Brownian motion. This additional information can change finite-horizon values but not κ H \kappa_H κ H . For any normalized horizon L > 0 L>0 L > 0 , let u H p a s t ( L ) u_H^{\mathrm{past}}(L) u H past ( L ) denote the optimal expected cost of the normalized cell problem with full past information and an additional terminal reset cost. Averaging over shifted partitions gives
0 ≤ u H p a s t ( L ) − κ H L ≤ 2 2 κ H , 0\leq u_H^{\mathrm{past}}(L)-\kappa_H L\leq2\sqrt{2\kappa_H}, 0 ≤ u H past ( L ) − κ H L ≤ 2 2 κ H ,
which also yields continuity in H H H when fractional Brownian motions with different Hurst parameters are constructed from the same Brownian motion. For nonstationary targets, let ρ : [ 0 , T ] → ( 0 , ∞ ) \rho:[0,T]\to(0,\infty) ρ : [ 0 , T ] → ( 0 , ∞ ) be the continuous local amplitude : near time t t t , target increments over a small interval of length h h h have characteristic scale ρ ( t ) h H \rho(t)h^H ρ ( t ) h H . Thus ρ ( t ) \rho(t) ρ ( t ) plays the local role of σ \sigma σ in the homogeneous model. The filtered tangent condition in assumption 6.1 couples the normalized local paths with standard fractional Brownian motion and matches their observation filtrations. Under this condition, for continuous positive weights a a a and b b b ,
λ − 2 H / ( 1 + H ) V λ ( Θ ) ⟶ λ → 0 κ H ∫ 0 T a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ρ ( t ) 2 / ( 1 + H ) d t . \lambda^{-2H/(1+H)}V_\lambda(\Theta)\mathop{\longrightarrow}_{\lambda\to0}\kappa_H\int_0^T a_t^{(1-H)/(1+H)}b_t^{2H/(1+H)}\rho(t)^{2/(1+H)}\,\mathrm{d}t. λ − 2 H / ( 1 + H ) V λ ( Θ ) ⟶ λ → 0 κ H ∫ 0 T a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ρ ( t ) 2/ ( 1 + H ) d t .
We verify the condition for Gaussian Volterra targets when their driving Brownian motion is observed. Perturbing the coefficients then identifies the leading tracking and variation costs for every asymptotically optimal family of controls , meaning one whose cost divided by V λ ( Θ ) V_\lambda(\Theta) V λ ( Θ ) tends to one.
Contributions. The asymptotic formula for fractional Brownian motion in theorem 3.4 extends to local Volterra targets through theorem 6.2 and proposition 6.8. The proof uses the information and reset estimates in theorems 5.3 and 5.4. Prediction states account for the past information carried by fractional targets. To estimate the cell constant, we derive primal and dual bounds and approximations based on exponential sums with positive weights in § 7 and § 8 .
Related work. Our Brownian benchmark builds on singular-control results [1 , 2 , 3 ] and small-cost tracking of continuous Itô targets [4 , 5 ] . Fractional linear–quadratic control with absolutely continuous controls is treated in [6 , 7 , 8 , 9 , 10 ] . Schelling [11 ] links rough-target tracking to an ergodic fractional cell under quadratic trading costs. In our setting, the penalty is linear in variation and permits singular controls. One-sided fractional singular control is studied in [12 , 13 ] . Truncated variation and hard-tube approximation [14 , 15 , 16 , 17 ] impose pathwise constraints. Related level-crossing asymptotics for fractional Brownian motion are studied in [18 ] . Prediction states and Markovian fractional or Volterra representations appear in [19 , 20 , 21 , 22 , 23 , 24 ] , while positive exponential approximations are studied in [25 , 26 , 27 , 28 ] . We approximate control values under a fixed observation filtration, with stationary-past and growing-horizon error bounds. The hedging application in section J gives a penalized mean-square identity. For self-financing models with proportional transaction costs in fractional finance, see [29 ] .
Organization. We first establish the rate and asymptotic formula in § 2 and § 3 , before turning to Brownian reflection and past information in § 4 and § 5 . These results lead to the local limit theorem and its Volterra application in § 6 . We then develop the bounds and factor approximations in § 7 and § 8 . The appendices give the filtration and path-space arguments, the prediction-state construction, and the technical proofs used in the main text. They also contain the multi-window bounds and the hedging identity in § 18 .§ 2 and J.
Model and rate bounds
Notation and admissibility
Fix a complete filtered probability space ( Ω , F , F , P ) (\Omega,\mathcal{F},\mathbf{F},\mathbb{P}) ( Ω , F , F , P ) , where F = ( F t ) t ≥ 0 \mathbf{F}=(\mathcal{F}_t)_{t\geq0} F = ( F t ) t ≥ 0 , and a horizon T > 0 T>0 T > 0 . All filtrations are complete and right-continuous. Expectations, variances, and covariances are denoted by E \mathbb{E} E , Var \operatorname{Var} Var , and Cov \operatorname{Cov} Cov . A bar denotes completion under the indicated probability measure. For a standard Borel space E E E , write B ( E ) \mathcal{B}(E) B ( E ) for its Borel sigma-field and P ( E ) \mathcal{P}(E) P ( E ) for the set of probability measures on E E E . We use L 0 ( Ω , H , P ; E ) L^0(\Omega,\mathcal{H},\mathbb{P};E) L 0 ( Ω , H , P ; E ) for the H \mathcal{H} H -measurable E E E -valued random variables, identified almost surely. For a separable Banach space ( E , ∥ ⋅ ∥ E ) (E,\|\cdot\|_E) ( E , ∥ ⋅ ∥ E ) and 1 ≤ p < ∞ 1\leq p<\infty 1 ≤ p < ∞ , the subspace L p ( Ω , H , P ; E ) L^p(\Omega,\mathcal{H},\mathbb{P};E) L p ( Ω , H , P ; E ) of L 0 ( Ω , H , P ; E ) L^0(\Omega,\mathcal{H},\mathbb{P};E) L 0 ( Ω , H , P ; E ) consists of the variables with E [ ∥ X ∥ E p ] < ∞ \mathbb{E}[\|X\|_E^p]<\infty E [ ∥ X ∥ E p ] < ∞ and carries the norm ∥ X ∥ L p : = ( E [ ∥ X ∥ E p ] ) 1 / p \|X\|_{L^p}:=(\mathbb{E}[\|X\|_E^p])^{1/p} ∥ X ∥ L p := ( E [ ∥ X ∥ E p ] ) 1/ p . Likewise, L ∞ L^\infty L ∞ uses the essential-supremum norm, while deterministic L p ( I ; E ) L^p(I;E) L p ( I ; E ) uses Lebesgue measure on I I I . Omitted sigma-fields and measures are those of the ambient space. An omitted scalar codomain is R \mathbb{R} R . E μ \mathbb{E}_\mu E μ denotes expectation under a specified law μ \mu μ . On R d \mathbb{R}^d R d , x ⋅ y = x ⊤ y x\cdot y=x^\top y x ⋅ y = x ⊤ y and ∥ x ∥ = ( x ⊤ x ) 1 / 2 \|x\|=(x^\top x)^{1/2} ∥ x ∥ = ( x ⊤ x ) 1/2 are the Euclidean inner product and norm. Inner products and norms with subscript 2 refer to the stated L 2 L^2 L 2 space. We write L ( X ) \mathcal{L}(X) L ( X ) for the law of X X X , = d \overset{d}{=} = d for equality in law, and U ⊥ ⊥ H ∣ Y U\perp\!\!\!\perp\mathcal{H}\mid Y U ⊥ ⊥ H ∣ Y for conditional independence given Y Y Y . Conditional essential infima are taken in the almost-sure order. C ( I ; E ) \mathcal{C}(I;E) C ( I ; E ) and D ( I ; E ) \mathcal{D}(I;E) D ( I ; E ) denote continuous and càdlàg paths, with the uniform and Skorokhod J 1 J_1 J 1 topologies on compact I I I . C l o c \mathcal{C}_{\mathrm{loc}} C loc uses uniform convergence on compact intervals. For Hölder estimates, use
[ f ] α : = sup s ≠ t ∣ f ( t ) − f ( s ) ∣ ∣ t − s ∣ α , 0 < α ≤ 1. [f]_\alpha:=\sup_{s\ne t}\frac{|f(t)-f(s)|}{|t-s|^\alpha},\qquad0<\alpha\leq1. [ f ] α := s = t sup ∣ t − s ∣ α ∣ f ( t ) − f ( s ) ∣ , 0 < α ≤ 1.
For 0 < α < 1 0<\alpha<1 0 < α < 1 , C α \mathcal{C}^\alpha C α is the space of continuous functions with finite [ f ] α [f]_\alpha [ f ] α . When derivatives are specified, C k \mathcal{C}^k C k denotes k k k continuous derivatives. Set C T : = C ( [ 0 , T ] ; R ) \mathcal{C}_T:=\mathcal{C}([0,T];\mathbb{R}) C T := C ([ 0 , T ] ; R ) and D T : = D ( [ 0 , T ] ; R ) \mathcal{D}_T:=\mathcal{D}([0,T];\mathbb{R}) D T := D ([ 0 , T ] ; R ) . For x ∈ D T x\in\mathcal{D}_T x ∈ D T and 0 ≤ r < s ≤ T 0\leq r<s\leq T 0 ≤ r < s ≤ T , define
TV ( r , s ] ( x ) : = sup r = t 0 < ⋯ < t n = s ∑ j = 1 n ∣ x t j − x t j − 1 ∣ . \operatorname{TV}_{(r,s]}(x):=\sup_{r=t_0<\cdots<t_n=s}\sum_{j=1}^{n}|x_{t_j}-x_{t_{j-1}}|. TV ( r , s ] ( x ) := r = t 0 < ⋯ < t n = s sup j = 1 ∑ n ∣ x t j − x t j − 1 ∣.
and let BV ( [ 0 , T ] ; R ) \operatorname{BV}([0,T];\mathbb{R}) BV ([ 0 , T ] ; R ) be the set of paths in D T \mathcal{D}_T D T with TV ( 0 , T ] ( x ) < ∞ \operatorname{TV}_{(0,T]}(x)<\infty TV ( 0 , T ] ( x ) < ∞ . Given a left value x r − x_{r-} x r − , include the initial adjustment by setting
TV [ r , s ] a u g ( x ) : = ∣ x r − x r − ∣ + TV ( r , s ] ( x ) . \operatorname{TV}^{\mathrm{aug}}_{[r,s]}(x):=|x_r-x_{r-}|+\operatorname{TV}_{(r,s]}(x). TV [ r , s ] aug ( x ) := ∣ x r − x r − ∣ + TV ( r , s ] ( x ) .
We write Δ x t : = x t − x t − \Delta x_t:=x_t-x_{t-} Δ x t := x t − x t − for the jump at time t t t . The signed Stieltjes measure d x dx d x includes the prescribed initial atom. Its total-variation measure ∣ d x ∣ |dx| ∣ d x ∣ is also written d ∣ x ∣ t d|x|_t d ∣ x ∣ t . The continuous part is x c x^c x c . R \mathbb{R} R , Q \mathbb{Q} Q , and N = { 1 , 2 , … } \mathbb{N}=\{1,2,\ldots\} N = { 1 , 2 , … } denote the real, rational, and positive integer numbers, and we set Q + : = Q ∩ [ 0 , ∞ ) \mathbb{Q}_+:=\mathbb{Q}\cap[0,\infty) Q + := Q ∩ [ 0 , ∞ ) . For real x , y x,y x , y , x + : = max ( x , 0 ) x^+:=\max(x,0) x + := max ( x , 0 ) , x ∧ y : = min ( x , y ) x\wedge y:=\min(x,y) x ∧ y := min ( x , y ) , x ∨ y : = max ( x , y ) x\vee y:=\max(x,y) x ∨ y := max ( x , y ) , and sign ( 0 ) : = 0 \operatorname{sign}(0):=0 sign ( 0 ) := 0 . We use 1 A \mathbf{1}_A 1 A for the indicator of A A A and ⌊ x ⌋ \lfloor x\rfloor ⌊ x ⌋ and ⌈ x ⌉ \lceil x\rceil ⌈ x ⌉ for the floor and ceiling. In linear algebra, span denotes the linear span, diag a diagonal matrix, and ⊤ \top ⊤ transpose. N ( 0 , 1 ) N(0,1) N ( 0 , 1 ) denotes a standard normal variable, and m 1 : = E [ ∣ N ( 0 , 1 ) ∣ ] = 2 / π m_1:=\mathbb{E}[|N(0,1)|]=\sqrt{2/\pi} m 1 := E [ ∣ N ( 0 , 1 ) ∣ ] = 2/ π . Constants C C C may change between lines. Subscripts indicate parameter dependence. The symbols O O O , o o o , ∼ \sim ∼ , and ≍ \asymp ≍ mean a bounded ratio, a vanishing ratio, a ratio tending to one, and two-sided comparison in the stated limit. Weak convergence of finite measures is tested against bounded continuous functions.
Definition 2.1 (augmented total variation and admissible controls). Let Θ ∈ L 0 ( Ω , F , P ; C T ) \Theta\in L^0(\Omega,\mathcal{F},\mathbb{P};\mathcal{C}_T) Θ ∈ L 0 ( Ω , F , P ; C T ) be F \mathbf{F} F -adapted. A control is a jointly F ⊗ B ( [ 0 , T ] ) \mathcal{F}\otimes\mathcal{B}([0,T]) F ⊗ B ([ 0 , T ]) -measurable map
ϕ : Ω × [ 0 , T ] ⟶ R ( ω , t ) ⟼ ϕ t ( ω ) \phi:\quad
\begin{array}{rcl}
\Omega\times[0,T]&\longrightarrow&\mathbb{R}\\
(\omega,t)&\longmapsto&\phi_t(\omega)
\end{array} ϕ : Ω × [ 0 , T ] ( ω , t ) ⟶ ⟼ R ϕ t ( ω ) such that
ϕ ( ω , ⋅ ) ∈ BV ( [ 0 , T ] ; R ) P -a.s. , ϕ t ∈ L 0 ( Ω , F t , P ) , 0 ≤ t ≤ T . \phi(\omega,\cdot)\in\operatorname{BV}([0,T];\mathbb{R})\quad\mathbb{P}\text{-a.s.},\qquad\phi_t\in L^0(\Omega,\mathcal{F}_t,\mathbb{P}),\quad0\leq t\leq T. ϕ ( ω , ⋅ ) ∈ BV ([ 0 , T ] ; R ) P -a.s. , ϕ t ∈ L 0 ( Ω , F t , P ) , 0 ≤ t ≤ T . With the pre-initial value ϕ 0 − : = Θ 0 \phi_{0-}:=\Theta_0 ϕ 0 − := Θ 0 , the augmented variation is
TV [ 0 , T ] a u g ( ϕ ) : = ∣ ϕ 0 − ϕ 0 − ∣ + TV ( 0 , T ] ( ϕ ) . (5) \operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi):=|\phi_0-\phi_{0-}|+\operatorname{TV}_{(0,T]}(\phi). \tag*{(5)} TV [ 0 , T ] aug ( ϕ ) := ∣ ϕ 0 − ϕ 0 − ∣ + TV ( 0 , T ] ( ϕ ) . (5) The measure ∣ d ϕ ∣ |d\phi| ∣ d ϕ ∣ on [ 0 , T ] [0,T] [ 0 , T ] is determined by
∣ d ϕ ∣ ( { 0 } ) = ∣ ϕ 0 − ϕ 0 − ∣ , ∣ d ϕ ∣ ( ( s , t ] ) = TV ( s , t ] ( ϕ ) , 0 ≤ s < t ≤ T . |d\phi|(\{0\})=|\phi_0-\phi_{0-}|,\qquad|d\phi|((s,t])=\operatorname{TV}_{(s,t]}(\phi),\quad0\leq s<t\leq T. ∣ d ϕ ∣ ({ 0 }) = ∣ ϕ 0 − ϕ 0 − ∣ , ∣ d ϕ ∣ (( s , t ]) = TV ( s , t ] ( ϕ ) , 0 ≤ s < t ≤ T . Let A T ( Θ ) \mathcal{A}_T(\Theta) A T ( Θ ) be the class of these controls satisfying
E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t + TV [ 0 , T ] a u g ( ϕ ) ] < ∞ . \mathbb{E}\left[\int_0^T(\phi_t-\Theta_t)^2\,dt+\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi)\right]<\infty. E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t + TV [ 0 , T ] aug ( ϕ ) ] < ∞. When Θ \Theta Θ is fixed, write A T : = A T ( Θ ) \mathcal{A}_T:=\mathcal{A}_T(\Theta) A T := A T ( Θ ) .
Definition 2.2 (tracking value). Let a , b : [ 0 , T ] → ( 0 , ∞ ) a,b:[0,T]\to(0,\infty) a , b : [ 0 , T ] → ( 0 , ∞ ) be Borel functions with a − ≤ a t ≤ a + a_{-}\le a_t\le a_{+} a − ≤ a t ≤ a + for d t \mathrm{d}t d t -a.e. t t t and b − ≤ b t ≤ b + b_{-}\le b_t\le b_{+} b − ≤ b t ≤ b + for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , where a − , b − > 0 a_{-},b_{-}>0 a − , b − > 0 and a + , b + < ∞ a_{+},b_{+}<\infty a + , b + < ∞ . For λ > 0 \lambda>0 λ > 0 , the tracking value is
V λ ( Θ ) : = inf ϕ ∈ A T ( Θ ) E [ 1 2 ∫ 0 T a t ( ϕ t − Θ t ) 2 d t + λ ∫ [ 0 , T ] b t ∣ d ϕ ∣ t ] . (6) V_{\lambda}(\Theta):=\inf_{\phi\in\mathcal{A}_{T}(\Theta)}\mathbb{E}\left[\frac{1}{2}\int_{0}^{T}a_t(\phi_t-\Theta_t)^2\mathrm{d}t+\lambda\int_{[0,T]}b_t\lvert\mathrm{d}\phi\rvert_t\right]. \tag*{(6)} V λ ( Θ ) := ϕ ∈ A T ( Θ ) inf E [ 2 1 ∫ 0 T a t ( ϕ t − Θ t ) 2 d t + λ ∫ [ 0 , T ] b t ∣ d ϕ ∣ t ] . (6) For x 0 ∈ R x_0\in\mathbb{R} x 0 ∈ R , V λ x 0 ( Θ ) V_{\lambda}^{x_0}(\Theta) V λ x 0 ( Θ ) denotes the same infimum with ϕ 0 − = x 0 \phi_{0-}=x_0 ϕ 0 − = x 0 .
Proposition 2.3 (effect of the initial position). Assume A T ( Θ ) ≠ ∅ \mathcal{A}_{T}(\Theta)\ne\varnothing A T ( Θ ) = ∅ and E [ ∣ Θ 0 ∣ ] < ∞ \mathbb{E}[\lvert\Theta_0\rvert]<\infty E [∣ Θ 0 ∣] < ∞ . Then
∣ V λ x 0 ( Θ ) − V λ ( Θ ) ∣ ≤ λ b + E [ ∣ x 0 − Θ 0 ∣ ] . (7) \left\lvert V_{\lambda}^{x_0}(\Theta)-V_{\lambda}(\Theta)\right\rvert\le\lambda b_{+}\mathbb{E}[\lvert x_0-\Theta_0\rvert]. \tag*{(7)} ∣ V λ x 0 ( Θ ) − V λ ( Θ ) ∣ ≤ λ b + E [∣ x 0 − Θ 0 ∣] . (7) Therefore, every asymptotic equivalent of order λ α \lambda^{\alpha} λ α with α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) is unchanged by a fixed initial position.
Proof. Write J λ x 0 J_{\lambda}^{x_0} J λ x 0 for the objective with pre-initial value x 0 x_0 x 0 . For the same control path, only the initial variation atom changes:
∣ J λ x 0 ( ϕ ; Θ ) − J λ ( ϕ ; Θ ) ∣ = λ b 0 ∣ E [ ∣ ϕ 0 − x 0 ∣ − ∣ ϕ 0 − Θ 0 ∣ ] ∣ ≤ λ b + E [ ∣ x 0 − Θ 0 ∣ ] . \begin{aligned}
\left\lvert J_{\lambda}^{x_0}(\phi;\Theta)-J_{\lambda}(\phi;\Theta)\right\rvert=\lambda b_0\left\lvert\mathbb{E}[\lvert\phi_0-x_0\rvert-\lvert\phi_0-\Theta_0\rvert]\right\rvert\\
&\le\lambda b_{+}\mathbb{E}[\lvert x_0-\Theta_0\rvert].
\end{aligned} ∣ J λ x 0 ( ϕ ; Θ ) − J λ ( ϕ ; Θ ) ∣ = λ b 0 ∣ E [∣ ϕ 0 − x 0 ∣ − ∣ ϕ 0 − Θ 0 ∣] ∣ ≤ λ b + E [∣ x 0 − Θ 0 ∣] .
By the triangle inequality, the two admissible classes contain the same paths. Taking infima in both directions proves (7) . The normalized error is bounded by b + E [ ∣ x 0 − Θ 0 ∣ ] λ 1 − α → λ → 0 0 b_{+}\mathbb{E}[\lvert x_0-\Theta_0\rvert]\lambda^{1-\alpha}\xrightarrow{\lambda\to0}0 b + E [∣ x 0 − Θ 0 ∣] λ 1 − α λ → 0 0 .
Proposition 2.4 (finite-variation perturbations). Assume A T ( Θ ) ≠ ∅ \mathcal{A}_{T}(\Theta)\ne\varnothing A T ( Θ ) = ∅ . Let A A A be continuous, adapted, and of finite variation, with E [ TV ( 0 , T ] ( A ) ] < ∞ \mathbb{E}[\operatorname{TV}_{(0,T]}(A)]<\infty E [ TV ( 0 , T ] ( A )] < ∞ . Using each target’s pre-initial value,
∣ V λ ( Θ + A ) − V λ ( Θ ) ∣ ≤ λ E [ ∫ ( 0 , T ] b t ∣ d A ∣ t ] . (8) \left\lvert V_{\lambda}(\Theta+A)-V_{\lambda}(\Theta)\right\rvert\le\lambda\mathbb{E}\left[\int_{(0,T]}b_t\lvert\mathrm{d}A\rvert_t\right]. \tag*{(8)} ∣ V λ ( Θ + A ) − V λ ( Θ ) ∣ ≤ λ E [ ∫ ( 0 , T ] b t ∣ d A ∣ t ] . (8) Thus an integrable finite-variation perturbation does not change a leading term of order λ α \lambda^{\alpha} λ α with 0 < α < 1 0<\alpha<1 0 < α < 1 .
Proof. Adding A A A defines a bijection from A T ( Θ ) \mathcal{A}_{T}(\Theta) A T ( Θ ) to A T ( Θ + A ) \mathcal{A}_{T}(\Theta+A) A T ( Θ + A ) . It preserves adaptedness and satisfies
( ϕ t + A t ) − ( Θ t + A t ) = ϕ t − Θ t , ∣ ( ϕ 0 + A 0 ) − ( Θ 0 + A 0 ) ∣ = ∣ ϕ 0 − Θ 0 ∣ . (\phi_t+A_t)-(\Theta_t+A_t)=\phi_t-\Theta_t,\qquad\lvert(\phi_{0}+A_{0})-(\Theta_{0}+A_{0})\rvert=\lvert\phi_{0}-\Theta_{0}\rvert. ( ϕ t + A t ) − ( Θ t + A t ) = ϕ t − Θ t , ∣( ϕ 0 + A 0 ) − ( Θ 0 + A 0 )∣ = ∣ ϕ 0 − Θ 0 ∣ .
For the signed Stieltjes measures on ( 0 , T ] (0,T] ( 0 , T ] ,
∣ d ( ϕ + A ) ∣ ≤ ∣ d ϕ ∣ + ∣ d A ∣ . \lvert\mathrm{d}(\phi+A)\rvert\le\lvert\mathrm{d}\phi\rvert+\lvert\mathrm{d}A\rvert. ∣ d ( ϕ + A )∣ ≤ ∣ d ϕ ∣ + ∣ d A ∣ .
Then,
∫ ( 0 , T ] b t ∣ d ( ϕ + A ) ∣ t ≤ ∫ ( 0 , T ] b t ∣ d ϕ ∣ t + ∫ ( 0 , T ] b t ∣ d A ∣ t . \int_{(0,T]}b_t\lvert\mathrm{d}(\phi+A)\rvert_t\le\int_{(0,T]}b_t\lvert\mathrm{d}\phi\rvert_t+\int_{(0,T]}b_t\lvert\mathrm{d}A\rvert_t. ∫ ( 0 , T ] b t ∣ d ( ϕ + A ) ∣ t ≤ ∫ ( 0 , T ] b t ∣ d ϕ ∣ t + ∫ ( 0 , T ] b t ∣ d A ∣ t .
Hence
V λ ( Θ + A ) ≤ V λ ( Θ ) + λ E [ ∫ ( 0 , T ] b t ∣ d A ∣ t ] . V_{\lambda}(\Theta+A)\le V_{\lambda}(\Theta)+\lambda\mathbb{E}\left[\int_{(0,T]}b_t\lvert\mathrm{d}A\rvert_t\right]. V λ ( Θ + A ) ≤ V λ ( Θ ) + λ E [ ∫ ( 0 , T ] b t ∣ d A ∣ t ] .
Replace ( Θ , A ) (\Theta,A) ( Θ , A ) by ( Θ + A , − A ) (\Theta+A,-A) ( Θ + A , − A ) for the reverse bound. Division by λ α \lambda^{\alpha} λ α leaves an O ( λ 1 − α ) O(\lambda^{1-\alpha}) O ( λ 1 − α ) error.
:::
Assumption 2.5 (two-sided increment scale). There are H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) , h 0 > 0 h_0>0 h 0 > 0 , and c 1 , C 2 > 0 c_1,C_2>0 c 1 , C 2 > 0 such that, for 0 < h < min ( h 0 , T ) 0<h<\min(h_0,T) 0 < h < min ( h 0 , T ) ,
∫ 0 T − h E [ ∣ Θ t + h − Θ t ∣ ] d t ≥ c 1 ( T − h ) h H , \int_{0}^{T-h}\mathbb{E}[\lvert\Theta_{t+h}-\Theta_t\rvert]\mathrm{d}t\ge c_1(T-h)h^{H}, ∫ 0 T − h E [∣ Θ t + h − Θ t ∣] d t ≥ c 1 ( T − h ) h H , and
sup 0 ≤ t ≤ T − h E [ ∣ Θ t + h − Θ t ∣ 2 ] ≤ C 2 h 2 H . (9) \sup_{0\le t\le T-h}\mathbb{E}[\lvert\Theta_{t+h}-\Theta_t\rvert^2]\le C_2h^{2H}. \tag*{(9)} 0 ≤ t ≤ T − h sup E [∣ Θ t + h − Θ t ∣ 2 ] ≤ C 2 h 2 H . (9)
Definition 2.6 (noncausal comparison class). Let A T a l l ( Θ ) \mathcal{A}^{\mathrm{all}}_{T}(\Theta) A T all ( Θ ) be the set of D T \mathcal{D}_{T} D T -valued random variables ϕ \phi ϕ satisfying the pathwise, pre-initial-value, and integrability conditions of A T ( Θ ) \mathcal{A}_{T}(\Theta) A T ( Θ ) , but without the conditions ϕ t ∈ L 0 ( Ω , F t , P ) \phi_t\in L^0(\Omega,\mathcal{F}_t,\mathbb{P}) ϕ t ∈ L 0 ( Ω , F t , P ) . When Θ \Theta Θ is fixed, write A T a l l : = A T a l l ( Θ ) \mathcal{A}^{\mathrm{all}}_{T}:=\mathcal{A}^{\mathrm{all}}_{T}(\Theta) A T all := A T all ( Θ ) .
Lemma 2.7 (increment versus energy). For ϕ ∈ A T a l l \phi\in\mathcal{A}^{\mathrm{all}}_{T} ϕ ∈ A T all , let M = E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t ] M=\mathbb{E}[\int_{0}^{T}(\phi_t-\Theta_t)^2\mathrm{d}t] M = E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t ] and R = E [ TV [ 0 , T ] a u g ( ϕ ) ] R=\mathbb{E}[\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi)] R = E [ TV [ 0 , T ] aug ( ϕ )] . For 0 < h < T 0<h<T 0 < h < T ,
∫ 0 T − h E [ ∣ Θ t + h − Θ t ∣ ] d t ≤ h R + 2 T M . (10) \int_{0}^{T-h}\mathbb{E}[\lvert\Theta_{t+h}-\Theta_t\rvert]\mathrm{d}t\le hR+2\sqrt{TM}. \tag*{(10)} ∫ 0 T − h E [∣ Θ t + h − Θ t ∣] d t ≤ h R + 2 T M . (10)
Proof. Put e : = ϕ − Θ ∈ L 2 ( Ω × [ 0 , T ] , P ⊗ d t ) e := \phi-\Theta\in L^2(\Omega\times[0,T], \mathbb{P}\otimes\mathrm{d}t) e := ϕ − Θ ∈ L 2 ( Ω × [ 0 , T ] , P ⊗ d t ) . Pathwise,
∣ Θ t + h − Θ t ∣ ≤ ∫ ( t , t + h ] ∣ d ϕ ∣ s + ∣ e t ∣ + ∣ e t + h ∣ . |\Theta_{t+h}-\Theta_t| \le\int_{(t,t+h]} |\mathrm{d}\phi|_s + |e_t| + |e_{t+h}|. ∣ Θ t + h − Θ t ∣ ≤ ∫ ( t , t + h ] ∣ d ϕ ∣ s + ∣ e t ∣ + ∣ e t + h ∣.
Tonelli’s theorem and the window length give
∫ 0 T − h ∫ ( t , t + h ] ∣ d ϕ ∣ s d t = ∫ ( 0 , T ] ( ∫ 0 T − h 1 { t < s ≤ t + h } d t ) ∣ d ϕ ∣ s = ∫ ( 0 , T ] ∣ [ s − h , s ) ∩ [ 0 , T − h ] ∣ ∣ d ϕ ∣ s ≤ h TV [ 0 , T ] a u g ( ϕ ) , \begin{aligned}
\int_0^{T-h}\int_{(t,t+h]}|\mathrm{d}\phi|_s\mathrm{d}t
= \int_{(0,T]}\left(\int_0^{T-h}\mathbf{1}_{\{t<s\le t+h\}}\mathrm{d}t\right)|\mathrm{d}\phi|_s \\
&= \int_{(0,T]} |[s-h,s)\cap[0,T-h]|\ |\mathrm{d}\phi|_s
\le h\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi),
\end{aligned} ∫ 0 T − h ∫ ( t , t + h ] ∣ d ϕ ∣ s d t = ∫ ( 0 , T ] ( ∫ 0 T − h 1 { t < s ≤ t + h } d t ) ∣ d ϕ ∣ s = ∫ ( 0 , T ] ∣ [ s − h , s ) ∩ [ 0 , T − h ] ∣ ∣ d ϕ ∣ s ≤ h TV [ 0 , T ] aug ( ϕ ) ,
where ∣ I ∣ |I| ∣ I ∣ denotes the Lebesgue length of an interval I I I . For either I = [ 0 , T − h ] I=[0,T-h] I = [ 0 , T − h ] or I = [ h , T ] I=[h,T] I = [ h , T ] , Cauchy–Schwarz yields
∫ I E [ ∣ e t ∣ ] d t ≤ ∣ I ∣ ( ∫ I E [ ∣ e t ∣ 2 ] d t ) 1 / 2 ≤ ∣ I ∣ ( ∫ I E [ e t 2 ] d t ) 1 / 2 ≤ T M . \begin{aligned}
\int_I \mathbb{E}[|e_t|]\mathrm{d}t
\le\sqrt{|I|}\left(\int_I\mathbb{E}[|e_t|^2]\mathrm{d}t\right)^{1/2} \\
&\le\sqrt{|I|}\left(\int_I\mathbb{E}[e_t^2]\mathrm{d}t\right)^{1/2}
\le\sqrt{TM}.
\end{aligned} ∫ I E [ ∣ e t ∣ ] d t ≤ ∣ I ∣ ( ∫ I E [ ∣ e t ∣ 2 ] d t ) 1/2 ≤ ∣ I ∣ ( ∫ I E [ e t 2 ] d t ) 1/2 ≤ T M .
Taking expectations and adding the two error bounds proves the claim.
Theorem 2.8 (rate bounds). Suppose assumption 2.5 holds. With the cutoff
λ 0 : = ( 1 ∧ h 0 2 ∧ T 2 ) 1 + H , (11) \lambda_0 := \left(1\wedge\frac{h_0}{2}\wedge\frac{T}{2}\right)^{1+H}, \tag*{(11)} λ 0 := ( 1 ∧ 2 h 0 ∧ 2 T ) 1 + H , (11) define the lower and upper constants by
C − : = T min { a − c 1 2 128 , b − c 1 4 } , (12) C_- := T\min\left\{\frac{a_-c_1^2}{128},\frac{b_-c_1}{4}\right\}, \tag*{(12)} C − := T min { 128 a − c 1 2 , 4 b − c 1 } , (12) and
C + : = a + C 2 T 2 + b + + C 2 ( T + 1 ) . (13) C_+ := \frac{a_+C_2T}{2}+b_++\sqrt{C_2}(T+1). \tag*{(13)} C + := 2 a + C 2 T + b + + C 2 ( T + 1 ) . (13) For 0 < λ ≤ λ 0 0<\lambda\le\lambda_0 0 < λ ≤ λ 0 , the value satisfies
C − λ 2 H / ( 1 + H ) ≤ V λ ( Θ ) ≤ C + λ 2 H / ( 1 + H ) . (14) C_-\lambda^{2H/(1+H)}\le V_\lambda(\Theta)\le C_+\lambda^{2H/(1+H)}. \tag*{(14)} C − λ 2 H / ( 1 + H ) ≤ V λ ( Θ ) ≤ C + λ 2 H / ( 1 + H ) . (14) The lower bound also holds when the infimum is taken over A T a l l \mathcal{A}^{\mathrm{all}}_T A T all .
Proof. Put h : = λ 1 / ( 1 + H ) ≤ 1 ∧ h 0 / 2 ∧ T / 2 h:=\lambda^{1/(1+H)}\le1\wedge h_0/2\wedge T/2 h := λ 1/ ( 1 + H ) ≤ 1 ∧ h 0 /2 ∧ T /2 and use M , R M,R M , R from lemma 2.7. For any admissible path, causal or not,
J λ ( ϕ ; Θ ) ≥ a − 2 M + λ b − R . J_\lambda(\phi;\Theta)\ge\frac{a_-}{2}M+\lambda b_-R. J λ ( ϕ ; Θ ) ≥ 2 a − M + λ b − R .
Moreover,
h R + 2 T M ≥ ∫ 0 T − h E [ ∣ Θ t + h − Θ t ∣ ] d t ≥ c 1 ( T − h ) h H ≥ c 1 T 2 h H . (15) \begin{aligned}
hR+2\sqrt{TM}\ge\int_0^{T-h}\mathbb{E}[|\Theta_{t+h}-\Theta_t|]\mathrm{d}t \\
&\ge c_1(T-h)h^H\ge\frac{c_1T}{2}h^H. \tag*{(15)}
\end{aligned} h R + 2 T M ≥ ∫ 0 T − h E [ ∣ Θ t + h − Θ t ∣ ] d t ≥ c 1 ( T − h ) h H ≥ 2 c 1 T h H . (15)
If M ≥ c 1 2 T h 2 H / 64 M\ge c_1^2Th^{2H}/64 M ≥ c 1 2 T h 2 H /64 , then
J λ ( ϕ ; Θ ) ≥ a − M 2 ≥ a − c 1 2 T 128 h 2 H . J_\lambda(\phi;\Theta)\ge\frac{a_-M}{2}\ge\frac{a_-c_1^2T}{128}h^{2H}. J λ ( ϕ ; Θ ) ≥ 2 a − M ≥ 128 a − c 1 2 T h 2 H .
Otherwise, the increment bound gives
R ≥ c 1 T 2 h H − 1 − 2 T M h ≥ c 1 T 4 h H − 1 , R\ge\frac{c_1T}{2}h^{H-1}-\frac{2\sqrt{TM}}{h}\ge\frac{c_1T}{4}h^{H-1}, R ≥ 2 c 1 T h H − 1 − h 2 T M ≥ 4 c 1 T h H − 1 ,
and therefore
J λ ( ϕ ; Θ ) ≥ λ b − R ≥ b − c 1 T 4 λ h H − 1 = b − c 1 T 4 h 2 H . J_\lambda(\phi;\Theta)\ge\lambda b_-R\ge\frac{b_-c_1T}{4}\lambda h^{H-1}=\frac{b_-c_1T}{4}h^{2H}. J λ ( ϕ ; Θ ) ≥ λ b − R ≥ 4 b − c 1 T λ h H − 1 = 4 b − c 1 T h 2 H .
Consequently,
J λ ( ϕ ; Θ ) ≥ T min { a − c 1 2 128 , b − c 1 4 } h 2 H = C − λ 2 H / ( 1 + H ) . (16) J_{\lambda}(\phi;\Theta) \ge T\min\left\{\frac{a-c_{1}^{2}}{128},\frac{b-c_{1}}{4}\right\}h^{2H}=C_{-}\lambda^{2H/(1+H)}. \tag*{(16)} J λ ( ϕ ; Θ ) ≥ T min { 128 a − c 1 2 , 4 b − c 1 } h 2 H = C − λ 2 H / ( 1 + H ) . (16)
For the upper bound, take a uniform mesh with step h h h and a possibly shorter last interval. Set N : = ⌈ T / h ⌉ N:=\lceil T/h\rceil N := ⌈ T / h ⌉ and t k : = min { k h , T } t_k:=\min\{kh,T\} t k := min { k h , T } for k = 0 , … , N k=0,\ldots,N k = 0 , … , N . On [ t k , t k + 1 ) [t_k,t_{k+1}) [ t k , t k + 1 ) , set ϕ t : = Θ t k \phi_t:=\Theta_{t_k} ϕ t := Θ t k , and choose ϕ T : = ϕ T − \phi_T:=\phi_{T-} ϕ T := ϕ T − to avoid a terminal jump. The increment bounds and Cauchy–Schwarz give
E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t ] ≤ C 2 ∑ k = 0 N − 1 ∫ t k t k + 1 ( t − t k ) 2 H d t ≤ C 2 T h 2 H . \mathbb{E}\left[\int_0^T(\phi_t-\Theta_t)^2\,\mathrm{d}t\right]\le C_2\sum_{k=0}^{N-1}\int_{t_k}^{t_{k+1}}(t-t_k)^{2H}\,\mathrm{d}t\le C_2Th^{2H}. E [ ∫ 0 T ( ϕ t − Θ t ) 2 d t ] ≤ C 2 k = 0 ∑ N − 1 ∫ t k t k + 1 ( t − t k ) 2 H d t ≤ C 2 T h 2 H .
and
E [ TV [ 0 , T ] a u g ( ϕ ) ] ≤ ∑ k = 1 N − 1 E [ ∣ Θ t k − Θ t k − 1 ∣ ] ≤ C 2 ( T / h + 1 ) h H . \mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi)\right]\le\sum_{k=1}^{N-1}\mathbb{E}\left[|\Theta_{t_k}-\Theta_{t_{k-1}}|\right]\le\sqrt{C_2(T/h+1)}h^H. E [ TV [ 0 , T ] aug ( ϕ ) ] ≤ k = 1 ∑ N − 1 E [ ∣ Θ t k − Θ t k − 1 ∣ ] ≤ C 2 ( T / h + 1 ) h H .
Therefore
J λ ( ϕ ; Θ ) ≤ a + C 2 T 2 h 2 H + λ b + C 2 ( T h H − 1 + h H ) = h 2 H [ a + C 2 T 2 + b + C 2 ( T + h ) ] ≤ C + λ 2 H / ( 1 + H ) . \begin{aligned}
J_{\lambda}(\phi;\Theta)\le\frac{a+C_2T}{2}h^{2H}+\lambda b+\sqrt{C_2}(Th^{H-1}+h^H) \\
&=h^{2H}\left[\frac{a+C_2T}{2}+b+\sqrt{C_2}(T+h)\right] \\
&\le C_{+}\lambda^{2H/(1+H)}.
\end{aligned} J λ ( ϕ ; Θ ) ≤ 2 a + C 2 T h 2 H + λb + C 2 ( T h H − 1 + h H ) = h 2 H [ 2 a + C 2 T + b + C 2 ( T + h ) ] ≤ C + λ 2 H / ( 1 + H ) .
The fractional Brownian cell and the asymptotic constant
Fix H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) . Standard fBm is the continuous centered Gaussian process B H B^H B H with B 0 H = 0 B^H_0=0 B 0 H = 0 and covariance
R H : ∣ R 2 ⟶ R ( s , t ) ⟼ 1 2 ( ∣ s ∣ 2 H + ∣ t ∣ 2 H − ∣ t − s ∣ 2 H ) (17) R_H:\left|\begin{aligned}
\mathbb{R}^{2}&\longrightarrow\mathbb{R}\\
(s,t)&\longmapsto\frac{1}{2}\left(|s|^{2H}+|t|^{2H}-|t-s|^{2H}\right)
\end{aligned}\right. \tag*{(17)} R H : R 2 ( s , t ) ⟶ R ⟼ 2 1 ( ∣ s ∣ 2 H + ∣ t ∣ 2 H − ∣ t − s ∣ 2 H ) (17)
In particular,
E [ ( B t H − B s H ) 2 ] = ∣ t − s ∣ 2 H , E [ B s H B t H ] = R H ( s , t ) , s , t ≥ 0. (18) \mathbb{E}\left[(B_t^H-B_s^H)^2\right]=|t-s|^{2H},\qquad\mathbb{E}\left[B_s^H B_t^H\right]=R_H(s,t),\qquad s,t\ge0. \tag*{(18)} E [ ( B t H − B s H ) 2 ] = ∣ t − s ∣ 2 H , E [ B s H B t H ] = R H ( s , t ) , s , t ≥ 0. (18)
For constant a , b , σ > 0 a,b,\sigma>0 a , b , σ > 0 , we retain the notation V λ ( T ; a , b , σ ) V_{\lambda}(T;a,b,\sigma) V λ ( T ; a , b , σ ) introduced in § 1 , using the usual natural filtration of B H B^H B H . On C L = C ( [ 0 , L ] ; R ) \mathcal{C}_L=\mathcal{C}([0,L];\mathbb{R}) C L = C ([ 0 , L ] ; R ) , let μ L H \mu_L^H μ L H be the fBm law and let F H , L \mathcal{F}^{H,L} F H , L be the usual augmented coordinate filtration. We use B H B^H B H for the coordinate process and D L = D ( [ 0 , L ] ; R ) \mathcal{D}_L=\mathcal{D}([0,L];\mathbb{R}) D L = D ([ 0 , L ] ; R ) for control paths.
Definition 3.1 (one-sided fractional cells). Let A L H \mathcal{A}_L^H A L H be the set of ψ ∈ L 0 ( C L , B ( C L ) , μ L H ; D L ) \psi\in L^0(\mathcal{C}_L,\mathcal{B}(\mathcal{C}_L),\mu_L^H;\mathcal{D}_L) ψ ∈ L 0 ( C L , B ( C L ) , μ L H ; D L ) that are F H , L \mathcal{F}^{H,L} F H , L -adapted, have pre-initial value ψ 0 − = 0 \psi_{0-}=0 ψ 0 − = 0 and finite variation TV [ 0 , L ] a u g ( ψ ) < ∞ \operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)<\infty TV [ 0 , L ] aug ( ψ ) < ∞ μ L H \mu_L^H μ L H -a.s., and satisfy the integrability condition
E [ ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] a u g ( ψ ) ] < ∞ . \mathbb{E}\left[\int_0^L(\psi_s-B_s^H)^2\,\mathrm{d}s+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)\right]<\infty. E [ ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] aug ( ψ ) ] < ∞.
Their tracking and reset values are, respectively,
v H ( L ) : = inf ψ ∈ A L H E [ 1 2 ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] a u g ( ψ ) ] , (19) v_H(L):=\inf_{\psi\in\mathcal{A}_L^H}\mathbb{E}\left[\frac{1}{2}\int_0^L(\psi_s-B_s^H)^2\,\mathrm{d}s+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)\right], \tag*{(19)} v H ( L ) := ψ ∈ A L H inf E [ 2 1 ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] aug ( ψ ) ] , (19)
and
u H ( L ) : = inf ψ ∈ A L H E [ 1 2 ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] a u g ( ψ ) + ∣ ψ L − B L H ∣ ] . (20) u_H(L):=\inf_{\psi\in\mathcal{A}_L^H}\mathbb{E}\left[\frac{1}{2}\int_0^L(\psi_s-B_s^H)^2\,\mathrm{d}s+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)+|\psi_L-B_L^H|\right]. \tag*{(20)} u H ( L ) := ψ ∈ A L H inf E [ 2 1 ∫ 0 L ( ψ s − B s H ) 2 d s + TV [ 0 , L ] aug ( ψ ) + ∣ ψ L − B L H ∣ ] . (20)
Proposition 3.2 (self-similar scaling identity). With ℓ λ \ell_{\lambda} ℓ λ from (3) ,
V λ ( T ; a , b , σ ) = a σ 2 ℓ λ 2 H + 1 v H ( T ℓ λ ) . (21) V_{\lambda}(T;a,b,\sigma)=a\sigma^2\ell_{\lambda}^{2H+1}v_H\left(\frac{T}{\ell_{\lambda}}\right). \tag*{(21)} V λ ( T ; a , b , σ ) = a σ 2 ℓ λ 2 H + 1 v H ( ℓ λ T ) . (21)
Equivalently, for L : = T / ℓ λ L:=T/\ell_{\lambda} L := T / ℓ λ ,
V λ ( T ; a , b , σ ) T a ( 1 − H ) / ( 1 + H ) ( λ b ) 2 H / ( 1 + H ) σ 2 / ( 1 + H ) = v H ( L ) L . (22) \frac{V_{\lambda}(T;a,b,\sigma)}{Ta^{(1-H)/(1+H)}(\lambda b)^{2H/(1+H)}\sigma^{2/(1+H)}}=\frac{v_H(L)}{L}. \tag*{(22)} T a ( 1 − H ) / ( 1 + H ) ( λb ) 2 H / ( 1 + H ) σ 2/ ( 1 + H ) V λ ( T ; a , b , σ ) = L v H ( L ) . (22)
Proof. For ℓ > 0 \ell> 0 ℓ > 0 , the continuous bijection
S ℓ : C T ⟶ C T / ℓ , x ⟼ ( s ↦ ℓ − H x ( ℓ s ) ) S_{\ell}:\quad\mathcal{C}_{T} \longrightarrow\mathcal{C}_{T/\ell},\qquad x \longmapsto(s \mapsto\ell^{-H}x(\ell s)) S ℓ : C T ⟶ C T / ℓ , x ⟼ ( s ↦ ℓ − H x ( ℓ s ))
transports fBm law and its natural filtration. Transport causal rules by theorem C.1. For 0 ≤ s ≤ T / ℓ 0 \le s \le T/\ell 0 ≤ s ≤ T / ℓ , rescale the target and control by
B ~ s H : = ℓ − H B ℓ s H , ψ s : = ϕ ℓ s σ ℓ H . \widetilde{B}^{H}_{s} := \ell^{-H}B^{H}_{\ell s},\qquad\psi_{s}:=\frac{\phi_{\ell s}}{\sigma\ell^{H}}. B s H := ℓ − H B ℓ s H , ψ s := σ ℓ H ϕ ℓ s .
The inverse is ϕ t = σ ℓ H ψ t / ℓ \phi_{t}=\sigma\ell^{H}\psi_{t/\ell} ϕ t = σ ℓ H ψ t / ℓ . Both maps preserve adaptedness, finite variation, and the zero pre-initial value. With t = ℓ s t=\ell s t = ℓ s ,
a 2 E [ ∫ 0 T ( ϕ t − σ B t H ) 2 d t ] = a σ 2 ℓ 2 H + 1 2 E [ ∫ 0 T / ℓ ( ψ s − B ~ s H ) 2 d s ] , \frac{a}{2}\mathbb{E}\left[\int_{0}^{T}(\phi_{t}-\sigma B^{H}_{t})^{2}\,\mathrm{d}t\right]=\frac{a\sigma^{2}\ell^{2H+1}}{2}\mathbb{E}\left[\int_{0}^{T/\ell}(\psi_{s}-\widetilde{B}^{H}_{s})^{2}\,\mathrm{d}s\right], 2 a E [ ∫ 0 T ( ϕ t − σ B t H ) 2 d t ] = 2 a σ 2 ℓ 2 H + 1 E [ ∫ 0 T / ℓ ( ψ s − B s H ) 2 d s ] ,
and
λ b E [ TV [ 0 , T ] a u g ( ϕ ) ] = λ b σ ℓ H E [ TV [ 0 , T / ℓ ] a u g ( ψ ) ] . \lambda b\mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\phi)\right]=\lambda b\sigma\ell^{H}\mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[0,T/\ell]}(\psi)\right]. λb E [ TV [ 0 , T ] aug ( ϕ ) ] = λbσ ℓ H E [ TV [ 0 , T / ℓ ] aug ( ψ ) ] .
Equality of the two coefficients is equivalent to
a σ 2 ℓ 2 H + 1 = λ b σ ℓ H ⟺ ℓ H + 1 = λ b a σ ⟺ ℓ = ℓ λ . a\sigma^{2}\ell^{2H+1}=\lambda b\sigma\ell^{H}\quad\Longleftrightarrow\quad\ell^{H+1}=\frac{\lambda b}{a\sigma}\quad\Longleftrightarrow\quad\ell=\ell_{\lambda}. a σ 2 ℓ 2 H + 1 = λbσ ℓ H ⟺ ℓ H + 1 = aσ λb ⟺ ℓ = ℓ λ .
Taking infima in both directions gives (21) . Finally,
a σ 2 ℓ λ 2 H = a σ 2 ( λ b a σ ) 2 H 1 + H = a 1 − H 1 + H ( λ b ) 2 H 1 + H σ 2 1 + H . a\sigma^{2}\ell_{\lambda}^{2H}=a\sigma^{2}\left(\frac{\lambda b}{a\sigma}\right)^{\frac{2H}{1+H}}=a^{\frac{1-H}{1+H}}(\lambda b)^{\frac{2H}{1+H}}\sigma^{\frac{2}{1+H}}. a σ 2 ℓ λ 2 H = a σ 2 ( aσ λb ) 1 + H 2 H = a 1 + H 1 − H ( λb ) 1 + H 2 H σ 1 + H 2 .
Divide (21) by T a σ 2 ℓ λ 2 H Ta\sigma^{2}\ell_{\lambda}^{2H} T a σ 2 ℓ λ 2 H to obtain (22) .
Proposition 3.3 (reset concatenation). For every S , T > 0 S,T>0 S , T > 0 ,
u H ( S + T ) ≤ u H ( S ) + u H ( T ) . (23) u_{H}(S+T)\le u_{H}(S)+u_{H}(T). \tag*{(23)} u H ( S + T ) ≤ u H ( S ) + u H ( T ) . (23)
Proof. Choose Borel reset rules ψ S , ψ T \psi^{S},\psi^{T} ψ S , ψ T with expected costs at most u H ( S ) + η u_{H}(S)+\eta u H ( S ) + η and u H ( T ) + η u_{H}(T)+\eta u H ( T ) + η , respectively. Apply the second rule to the increment path Y t : = B S + t H − B S H Y_{t}:=B^{H}_{S+t}-B^{H}_{S} Y t := B S + t H − B S H , 0 ≤ t ≤ T 0\le t\le T 0 ≤ t ≤ T , and concatenate by
ψ t : = { ψ t S , 0 ≤ t < S , B S H + ψ t − S T ( Y ) , S ≤ t ≤ S + T . \psi_{t}:=
\begin{cases}
\psi^{S}_{t}, & 0\le t<S,\\
B^{H}_{S}+\psi^{T}_{t-S}(Y), & S\le t\le S+T.
\end{cases} ψ t := { ψ t S , B S H + ψ t − S T ( Y ) , 0 ≤ t < S , S ≤ t ≤ S + T .
Stationary increments give L ( Y ) = μ H T \mathcal{L}(Y)=\mu^{T}_{H} L ( Y ) = μ H T , and the second rule is adapted to the original filtration. At the interface,
∣ B S H + ψ 0 T ( Y ) − ψ S − S ∣ ≤ ∣ Δ ψ S S ∣ + ∣ ψ S S − B S H ∣ + ∣ ψ 0 T ( Y ) ∣ . \left|B^{H}_{S}+\psi^{T}_{0}(Y)-\psi^{S}_{S-}\right|\le\left|\Delta\psi^{S}_{S}\right|+\left|\psi^{S}_{S}-B^{H}_{S}\right|+\left|\psi^{T}_{0}(Y)\right|. B S H + ψ 0 T ( Y ) − ψ S − S ≤ Δ ψ S S + ψ S S − B S H + ψ 0 T ( Y ) .
The first rule’s variation, its reset cost, and the second rule’s initial adjustment cost cover these terms, respectively. Running costs add, and ∣ ψ S + T − B S + T H ∣ = ∣ ψ T T ( Y ) − Y T ∣ \left|\psi_{S+T}-B^{H}_{S+T}\right|=\left|\psi^{T}_{T}(Y)-Y_{T}\right| ψ S + T − B S + T H = ψ T T ( Y ) − Y T . Thus lemma C.3 gives
u H ( S + T ) ≤ u H ( S ) + u H ( T ) + 2 η . u_{H}(S+T)\le u_{H}(S)+u_{H}(T)+2\eta. u H ( S + T ) ≤ u H ( S ) + u H ( T ) + 2 η .
Letting η → 0 \eta\to0 η → 0 proves the claim. Figure 1 summarizes the construction.
Figure 1. Reset concatenation of two causal cell rules. The second rule tracks the increment path Y r = B S + r H − B S H Y_r = B^H_{S+r} - B^H_S Y r = B S + r H − B S H and is translated by B S H B^H_S B S H . The virtual reset and the second rule’s initial adjustment bound the interface variation.
Theorem 3.4 (cell limit and small-cost asymptotic). For every H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) , there is a constant κ H ∈ ( 0 , ∞ ) \kappa_{H}\in(0,\infty) κ H ∈ ( 0 , ∞ ) such that
lim L → ∞ v H ( L ) L = lim L → ∞ u H ( L ) L = κ H , κ H = inf L > 0 u H ( L ) L . (24) \lim_{L\to\infty}\frac{v_{H}(L)}{L}=\lim_{L\to\infty}\frac{u_{H}(L)}{L}=\kappa_{H},\qquad\kappa_{H}=\inf_{L>0}\frac{u_{H}(L)}{L}. \tag*{(24)} L → ∞ lim L v H ( L ) = L → ∞ lim L u H ( L ) = κ H , κ H = L > 0 inf L u H ( L ) . (24) Therefore,
V λ ( T ; a , b , σ ) ∼ λ → 0 κ H T a 1 − H 1 + H ( λ b ) 2 H 1 + H σ 2 1 + H . (25) V_{\lambda}(T;a,b,\sigma)\mathop{\sim}_{\lambda\to0}\kappa_{H}Ta^{\frac{1-H}{1+H}}(\lambda b)^{\frac{2H}{1+H}}\sigma^{\frac{2}{1+H}}. \tag*{(25)} V λ ( T ; a , b , σ ) ∼ λ → 0 κ H T a 1 + H 1 − H ( λb ) 1 + H 2 H σ 1 + H 2 . (25)
Proof. Reset-cell limit. The zero rule gives
u H ( L ) ≤ 1 2 ∫ 0 L s 2 H d s + E [ ∣ B L H ∣ ] = L 2 H + 1 2 ( 2 H + 1 ) + m 1 L H . u_{H}(L)\le\frac{1}{2}\int_{0}^{L}s^{2H}\,\mathrm{d}s+\mathbb{E}\left[|B^{H}_{L}|\right]=\frac{L^{2H+1}}{2(2H+1)}+m_{1}L^{H}. u H ( L ) ≤ 2 1 ∫ 0 L s 2 H d s + E [ ∣ B L H ∣ ] = 2 ( 2 H + 1 ) L 2 H + 1 + m 1 L H .
Hence u H u_H u H is locally bounded. Subadditivity and section E yield
lim L → ∞ u H ( L ) L = inf L > 0 u H ( L ) L = : κ H . (26) \lim_{L\to\infty} \frac{u_H(L)}{L} = \inf_{L>0} \frac{u_H(L)}{L} =: \kappa_H. \tag*{(26)} L → ∞ lim L u H ( L ) = L > 0 inf L u H ( L ) =: κ H . (26)
A rule that samples the target at integer times and holds its value between samples, followed by a reset, satisfies
1 2 E [ ∫ 0 L ( ψ s − B s H ) 2 d s ] ≤ L / 2 , \frac{1}{2}\mathbb{E}\left[\int_0^L (\psi_s-B^H_s)^2\,\mathrm{d}s\right]\le L/2, 2 1 E [ ∫ 0 L ( ψ s − B s H ) 2 d s ] ≤ L /2 ,
and
E [ TV [ 0 , L ] a u g ( ψ ) + ∣ ψ L − B L H ∣ ] ≤ m 1 ( L + 1 ) , \mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)+|\psi_L-B^H_L|\right]\le m_1(L+1), E [ TV [ 0 , L ] aug ( ψ ) + ∣ ψ L − B L H ∣ ] ≤ m 1 ( L + 1 ) ,
so v H ( L ) ≤ u H ( L ) = O ( 1 + L ) v_H(L)\le u_H(L)=O(1+L) v H ( L ) ≤ u H ( L ) = O ( 1 + L ) .
Removing the reset. For L ≥ 1 L\ge1 L ≥ 1 , take ψ \psi ψ with non-reset cost at most v H ( L ) + η v_H(L)+\eta v H ( L ) + η . Since
∫ L − 1 L E [ ∣ ψ r − B r H ∣ 2 ] d r ≤ 2 ( v H ( L ) + η ) , \int_{L-1}^{L}\mathbb{E}\left[|\psi_r-B^H_r|^2\right]\,\mathrm{d}r\le2(v_H(L)+\eta), ∫ L − 1 L E [ ∣ ψ r − B r H ∣ 2 ] d r ≤ 2 ( v H ( L ) + η ) ,
choose a deterministic r ∈ [ L − 1 , L ) r\in[L-1,L) r ∈ [ L − 1 , L ) with E [ ∣ ψ r − B r H ∣ 2 ] ≤ 2 ( v H ( L ) + η ) \mathbb{E}[|\psi_r-B^H_r|^2]\le2(v_H(L)+\eta) E [ ∣ ψ r − B r H ∣ 2 ] ≤ 2 ( v H ( L ) + η ) . Keep ψ \psi ψ before r r r , jump to B r H B^H_r B r H , and hold until L L L . The interface inequality
∣ B r H − ψ r − ∣ ≤ ∣ Δ ψ r ∣ + ∣ ψ r − B r H ∣ |B^H_r-\psi_{r-}|\le|\Delta\psi_r|+|\psi_r-B^H_r| ∣ B r H − ψ r − ∣ ≤ ∣Δ ψ r ∣ + ∣ ψ r − B r H ∣
shows that the extra variation is at most ∣ ψ r − B r H ∣ |\psi_r-B^H_r| ∣ ψ r − B r H ∣ . The other added terms satisfy
1 2 E [ ∫ r L ( B t H − B r H ) 2 d t ] = ( L − r ) 2 H + 1 2 ( 2 H + 1 ) ≤ 1 2 ( 2 H + 1 ) , \frac{1}{2}\mathbb{E}\left[\int_r^L(B^H_t-B^H_r)^2\,\mathrm{d}t\right]
=\frac{(L-r)^{2H+1}}{2(2H+1)}
\le\frac{1}{2(2H+1)}, 2 1 E [ ∫ r L ( B t H − B r H ) 2 d t ] = 2 ( 2 H + 1 ) ( L − r ) 2 H + 1 ≤ 2 ( 2 H + 1 ) 1 ,
and
E [ ∣ B L H − B r H ∣ ] = m 1 ( L − r ) H ≤ m 1 . \mathbb{E}[|B^H_L-B^H_r|]=m_1(L-r)^H\le m_1. E [ ∣ B L H − B r H ∣ ] = m 1 ( L − r ) H ≤ m 1 .
Cauchy–Schwarz and η → 0 \eta\to0 η → 0 therefore give
0 ≤ u H ( L ) − v H ( L ) ≤ 2 v H ( L ) + 1 2 ( 2 H + 1 ) + m 1 , L ≥ 1. (27) 0\le u_H(L)-v_H(L)\le\sqrt{2v_H(L)}+\frac{1}{2(2H+1)}+m_1,\qquad L\ge1. \tag*{(27)} 0 ≤ u H ( L ) − v H ( L ) ≤ 2 v H ( L ) + 2 ( 2 H + 1 ) 1 + m 1 , L ≥ 1. (27)
The gap is O ( L ) O(\sqrt{L}) O ( L ) , so v H ( L ) / L → L → ∞ κ H v_H(L)/L\xrightarrow{L\to\infty}\kappa_H v H ( L ) / L L → ∞ κ H .
Positivity and scaling. For standard fBm on [ 0 , 1 ] [0,1] [ 0 , 1 ] , the increment assumptions hold with c 1 = m 1 c_1=m_1 c 1 = m 1 and C 2 = 1 C_2=1 C 2 = 1 . Taking λ = L − ( 1 + H ) \lambda=L^{-(1+H)} λ = L − ( 1 + H ) gives ℓ λ = L − 1 \ell_\lambda=L^{-1} ℓ λ = L − 1 and, for large L L L ,
L − ( 2 H + 1 ) v H ( L ) = V λ ( 1 ; 1 , 1 , 1 ) ≥ C λ 2 H / ( 1 + H ) = C L − 2 H . L^{-(2H+1)}v_H(L)=V_\lambda(1;1,1,1)\ge C\lambda^{2H/(1+H)}=CL^{-2H}. L − ( 2 H + 1 ) v H ( L ) = V λ ( 1 ; 1 , 1 , 1 ) ≥ C λ 2 H / ( 1 + H ) = C L − 2 H .
Thus v H ( L ) / L ≥ C > 0 v_H(L)/L\ge C>0 v H ( L ) / L ≥ C > 0 . Now use (22) , since T / ℓ λ → λ → 0 ∞ T/\ell_\lambda\xrightarrow{\lambda\to0}\infty T / ℓ λ λ → 0 ∞ .
Corollary 3.5 (sublinear reset gap). There is C H < ∞ C_H<\infty C H < ∞ such that
0 ≤ u H ( L ) − v H ( L ) ≤ C H ( 1 + L ) , L ≥ 1. (28) 0\le u_H(L)-v_H(L)\le C_H(1+\sqrt{L}),\qquad L\ge1. \tag*{(28)} 0 ≤ u H ( L ) − v H ( L ) ≤ C H ( 1 + L ) , L ≥ 1. (28) The reset-gap estimate follows from (26) and the upper bound obtained with a time step of one.
The Brownian case
For H = 1 / 2 H = 1/2 H = 1/2 , the tracking error is a sufficient state for the controlled Brownian dynamics.
Proposition 4.1 (Brownian cell constant). Reflection of e = ψ − B e = \psi- B e = ψ − B on [ − d ∗ , d ∗ ] [-d_\ast, d_\ast] [ − d ∗ , d ∗ ] is optimal in the long-run cell, where
d ∗ = ( 3 / 2 ) 1 / 3 , κ 1 / 2 = d ∗ 2 2 = 1 2 ( 3 / 2 ) 2 / 3 . (29) d_\ast= (3/2)^{1/3}, \qquad\kappa_{1/2} = \frac{d_\ast^2}{2} = \frac{1}{2}(3/2)^{2/3}. \tag*{(29)} d ∗ = ( 3/2 ) 1/3 , κ 1/2 = 2 d ∗ 2 = 2 1 ( 3/2 ) 2/3 . (29) For reflection on [ − d , d ] [-d,d] [ − d , d ] , the long-run tracking and variation cost rates are d 2 / 6 d^2/6 d 2 /6 and 1 / ( 2 d ) 1/(2d) 1/ ( 2 d ) , respectively. The total cost rate is
c : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) , d ⟼ d 2 6 + 1 2 d . c:\quad(0,\infty) \longrightarrow(0,\infty), \qquad d \longmapsto\frac{d^2}{6}+\frac{1}{2d}. c : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) , d ⟼ 6 d 2 + 2 d 1 . Its unique minimum occurs at d 3 = 3 / 2 d^3 = 3/2 d 3 = 3/2 .
Proof. See section F . □
Past information and prediction states
An interior cell has access to observations made before its local starting time. We show that this information does not change the long-run value.
The cell with past information
Let B H = ( B t H ) t ∈ R B^H = (B_t^H)_{t\in\mathbb{R}} B H = ( B t H ) t ∈ R be a two-sided standard fBm with B 0 H = 0 B_0^H = 0 B 0 H = 0 and covariance R H R_H R H . Its canonical space is C 0 ( R ) = { x ∈ C ( R ; R ) : x ( 0 ) = 0 } \mathcal{C}_0(\mathbb{R}) = \{x \in\mathcal{C}(\mathbb{R};\mathbb{R}) : x(0) = 0\} C 0 ( R ) = { x ∈ C ( R ; R ) : x ( 0 ) = 0 } , with law μ H p a s t \mu_H^{\mathrm{past}} μ H past and the local-uniform topology. For control times t ≥ 0 t \ge0 t ≥ 0 , use the full-past filtration F t B , p a s t : = σ ( B s H : s ≤ t ) μ H p a s t \mathcal{F}_t^{B,\mathrm{past}} := \sigma(B_s^H : s \le t)^{\mu_H^{\mathrm{past}}} F t B , past := σ ( B s H : s ≤ t ) μ H past , which is right-continuous by lemma A.1 .
Definition 5.1 (cells with past information). Let A L H , p a s t \mathcal{A}_L^{H,\mathrm{past}} A L H , past be the admissible class on [ 0 , L ] [0,L] [ 0 , L ] with pre-initial value zero and filtration F B \mathcal{F}^B F B . Its tracking and reset values are, respectively,
v H p a s t ( L ) : = inf ψ ∈ A L H , p a s t E [ 1 2 ∫ 0 L ( ψ t − B t H ) 2 d t + TV [ 0 , L ] a u g ( ψ ) ] , (30) v_H^{\mathrm{past}}(L) := \inf_{\psi\in\mathcal{A}_L^{H,\mathrm{past}}} \mathbb{E}\left[\frac{1}{2}\int_0^L(\psi_t-B_t^H)^2\,\mathrm{d}t+\operatorname{TV}_{[0,L]}^{\mathrm{aug}}(\psi)\right], \tag*{(30)} v H past ( L ) := ψ ∈ A L H , past inf E [ 2 1 ∫ 0 L ( ψ t − B t H ) 2 d t + TV [ 0 , L ] aug ( ψ ) ] , (30) and
u H p a s t ( L ) : = inf ψ ∈ A L H , p a s t E [ 1 2 ∫ 0 L ( ψ t − B t H ) 2 d t + TV [ 0 , L ] a u g ( ψ ) + ∣ ψ L − B L H ∣ ] . (31) u_H^{\mathrm{past}}(L) := \inf_{\psi\in\mathcal{A}_L^{H,\mathrm{past}}} \mathbb{E}\left[\frac{1}{2}\int_0^L(\psi_t-B_t^H)^2\,\mathrm{d}t+\operatorname{TV}_{[0,L]}^{\mathrm{aug}}(\psi)+|\psi_L-B_L^H|\right]. \tag*{(31)} u H past ( L ) := ψ ∈ A L H , past inf E [ 2 1 ∫ 0 L ( ψ t − B t H ) 2 d t + TV [ 0 , L ] aug ( ψ ) + ∣ ψ L − B L H ∣ ] . (31)
For M > 0 M>0 M > 0 , let F ( M ) \mathcal{F}^{(M)} F ( M ) be the usual augmentation of σ ( B s H : − M ≤ s ≤ t ) \sigma(B_s^H:-M\le s\le t) σ ( B s H : − M ≤ s ≤ t ) , 0 ≤ t ≤ L 0\le t\le L 0 ≤ t ≤ L . Restricting controls to this filtration defines u H , M p a s t ( L ) u_{H,M}^{\mathrm{past}}(L) u H , M past ( L ) and v H , M p a s t ( L ) v_{H,M}^{\mathrm{past}}(L) v H , M past ( L ) . Write A L , M H , p a s t \mathcal{A}_{L,M}^{H,\mathrm{past}} A L , M H , past for the corresponding admissible class.
Lemma 5.2 (finite-memory approximation). For every fixed L > 0 L>0 L > 0 ,
u H , M p a s t ( L ) ↓ M → ∞ u H p a s t ( L ) . (32) u_{H,M}^{\mathrm{past}}(L) \downarrow_{M\to\infty} u_H^{\mathrm{past}}(L). \tag*{(32)} u H , M past ( L ) ↓ M → ∞ u H past ( L ) . (32) The same result holds for v H p a s t ( L ) v_H^{\mathrm{past}}(L) v H past ( L ) .
Proof. See section D . □
Theorem 5.3 (effect of past information). For every H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) ,
lim L → ∞ v H p a s t ( L ) L = lim L → ∞ u H p a s t ( L ) L = inf L > 0 u H p a s t ( L ) L = κ H . (33) \lim_{L\to\infty}\frac{v_H^{\mathrm{past}}(L)}{L}
=
\lim_{L\to\infty}\frac{u_H^{\mathrm{past}}(L)}{L}
=
\inf_{L>0}\frac{u_H^{\mathrm{past}}(L)}{L}
=
\kappa_H. \tag*{(33)} L → ∞ lim L v H past ( L ) = L → ∞ lim L u H past ( L ) = L > 0 inf L u H past ( L ) = κ H . (33)
Proof. The process ( B S + t H − B S H ) t ∈ R (B_{S+t}^H-B_S^H)_{t\in\mathbb{R}} ( B S + t H − B S H ) t ∈ R has the two-sided fBm law and is observed up to index t t t at time S + t S+t S + t . Reset concatenation gives
u H p a s t ( S + T ) ≤ u H p a s t ( S ) + u H p a s t ( T ) . u_H^{\mathrm{past}}(S+T)\le u_H^{\mathrm{past}}(S)+u_H^{\mathrm{past}}(T). u H past ( S + T ) ≤ u H past ( S ) + u H past ( T ) .
The zero rule and the rule with time step one give local boundedness and an O ( 1 + L ) O(1+L) O ( 1 + L ) upper bound. The reset-removal proof of theorem 3.4 uses only increment moments and yields
0 ≤ u H p a s t ( L ) − v H p a s t ( L ) ≤ 2 v H p a s t ( L ) + 1 2 ( 2 H + 1 ) + m 1 . 0\le u_H^{\mathrm{past}}(L)-v_H^{\mathrm{past}}(L)
\le\sqrt{2v_H^{\mathrm{past}}(L)}+\frac{1}{2(2H+1)}+m_1. 0 ≤ u H past ( L ) − v H past ( L ) ≤ 2 v H past ( L ) + 2 ( 2 H + 1 ) 1 + m 1 .
Thus both normalized values have the common limit
inf L > 0 u H p a s t ( L ) L ≤ κ H , \inf_{L>0}\frac{u^{\mathrm{past}}_H(L)}{L}\le\kappa_H, L > 0 inf L u H past ( L ) ≤ κ H ,
where the inequality follows by ignoring the past. For the converse, fix L , M > 0 L,M>0 L , M > 0 and a finite-memory reset rule of cost at most u H , M p a s t ( L ) + η u^{\mathrm{past}}_{H,M}(L)+\eta u H , M past ( L ) + η . On a one-sided path, first run a finite-cost reset rule on [ 0 , M ] [0,M] [ 0 , M ] , with cost C M C_M C M . On each block starting at s k : = M + k L s_k:=M+kL s k := M + k L , apply the selected rule to the segment
( B s k + t H − B s k H ) − M ≤ t ≤ L , \left(B^H_{s_k+t}-B^H_{s_k}\right)_{-M\le t\le L}, ( B s k + t H − B s k H ) − M ≤ t ≤ L ,
and add B s k H B^H_{s_k} B s k H to its output. This history is observed because s k − M ≥ 0 s_k-M\ge0 s k − M ≥ 0 , and stationary increments give the required segment law. The reset costs cover the variation at every interface, hence
u H ( M + n L ) ≤ C M + n { u H , M p a s t ( L ) + η } . u_H(M+nL)\le C_M+n\{u^{\mathrm{past}}_{H,M}(L)+\eta\}. u H ( M + n L ) ≤ C M + n { u H , M past ( L ) + η } .
After division by M + n L M+nL M + n L and n → ∞ n\to\infty n → ∞ ,
κ H ≤ u H , M p a s t ( L ) + η L → η → 0 , M → ∞ u H p a s t ( L ) L \kappa_H\le\frac{u^{\mathrm{past}}_{H,M}(L)+\eta}{L}\xrightarrow[\eta\to0,\ M\to\infty]{}\frac{u^{\mathrm{past}}_H(L)}{L} κ H ≤ L u H , M past ( L ) + η η → 0 , M → ∞ L u H past ( L )
by lemma 5.2. Taking the infimum over L L L proves equality.
Theorem 5.4 (a bounded full-past reset correction). For every H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) and L > 0 L>0 L > 0 ,
0 ≤ u H p a s t ( L ) − κ H L ≤ 2 2 κ H . (34) 0\le u^{\mathrm{past}}_H(L)-\kappa_H L\le2\sqrt{2\kappa_H}. \tag*{(34)} 0 ≤ u H past ( L ) − κ H L ≤ 2 2 κ H . (34) In particular,
0 ≤ u H p a s t ( L ) L − κ H ≤ C 0 L , with C 0 : = 2 2 ( 1 2 + 2 π ) , 0\le\frac{u^{\mathrm{past}}_H(L)}{L}-\kappa_H\le\frac{C_0}{L},\qquad\text{with }C_0:=2\sqrt{2\left(\frac12+\frac{\sqrt{2}}{\pi}\right)}, 0 ≤ L u H past ( L ) − κ H ≤ L C 0 , with C 0 := 2 2 ( 2 1 + π 2 ) , uniformly in H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) .
Proof. The lower bound is (33) . Fix L > 0 L>0 L > 0 . For R > 2 L R>2L R > 2 L , take a past-adapted rule with non-reset cost J R J_R J R satisfying
J R R → R → ∞ κ H . \frac{J_R}{R}\xrightarrow[R\to\infty]{}\kappa_H. R J R R → ∞ κ H .
For the error e : = ψ − B H e:=\psi-B^H e := ψ − B H , the energy M R : = E [ ∫ 0 R e t 2 d t ] M_R:=\mathbb{E}\left[\int_0^R e_t^2\,\mathrm{d}t\right] M R := E [ ∫ 0 R e t 2 d t ] therefore satisfies M R ≤ 2 J R M_R\le2J_R M R ≤ 2 J R . For a deterministic s ∈ [ 0 , L ) s\in[0,L) s ∈ [ 0 , L ) , set n s : = ⌊ ( R − s ) / L ⌋ n_s:=\lfloor(R-s)/L\rfloor n s := ⌊( R − s ) / L ⌋ and use grid points s + j L s+jL s + j L , 0 ≤ j ≤ n s 0\le j\le n_s 0 ≤ j ≤ n s . Each complete block, translated and given a virtual initial adjustment and terminal reset, is a full-past cell. Disjointness of the original costs and lemma H.1 imply, for almost every s s s ,
n s u H p a s t ( L ) ≤ J R + 2 ∑ j = 0 n s E [ ∣ e s + j L ∣ ] , n s ≥ R / L − 2. n_su^{\mathrm{past}}_H(L)\le J_R+2\sum_{j=0}^{n_s}\mathbb{E}\left[|e_{s+jL}|\right],\qquad n_s\ge R/L-2. n s u H past ( L ) ≤ J R + 2 j = 0 ∑ n s E [ ∣ e s + j L ∣ ] , n s ≥ R / L − 2.
Averaging over the grid offset and applying Tonelli’s theorem gives
∫ 0 L ∑ j = 0 ⌊ ( R − s ) / L ⌋ E [ ∣ e s + j L ∣ ] d s = ∑ j = 0 ∞ ∫ 0 L 1 { s + j L ≤ R } E [ ∣ e s + j L ∣ ] d s = ∑ j = 0 ∞ ∫ [ j L , ( j + 1 ) L ) ∩ [ 0 , R ] E [ ∣ e t ∣ ] d t = ∫ 0 R E [ ∣ e t ∣ ] d t . \begin{aligned}
\int_0^L\sum_{j=0}^{\lfloor(R-s)/L\rfloor}\mathbb{E}\left[|e_{s+jL}|\right]\,\mathrm{d}s
=\sum_{j=0}^{\infty}\int_0^L\mathbf{1}_{\{s+jL\le R\}}\mathbb{E}\left[|e_{s+jL}|\right]\,\mathrm{d}s \\
&=\sum_{j=0}^{\infty}\int_{[jL,(j+1)L)\cap[0,R]}\mathbb{E}\left[|e_t|\right]\,\mathrm{d}t
=\int_0^R\mathbb{E}\left[|e_t|\right]\,\mathrm{d}t.
\end{aligned} ∫ 0 L j = 0 ∑ ⌊( R − s ) / L ⌋ E [ ∣ e s + j L ∣ ] d s = j = 0 ∑ ∞ ∫ 0 L 1 { s + j L ≤ R } E [ ∣ e s + j L ∣ ] d s = j = 0 ∑ ∞ ∫ [ j L , ( j + 1 ) L ) ∩ [ 0 , R ] E [ ∣ e t ∣ ] d t = ∫ 0 R E [ ∣ e t ∣ ] d t .
Thus
( R L − 2 ) u H p a s t ( L ) ≤ J R + 2 L ∫ 0 R E [ ∣ e t ∣ ] d t ≤ J R + 2 L R M R . \left(\frac{R}{L}-2\right)u^{\mathrm{past}}_H(L)\le J_R+\frac{2}{L}\int_0^R\mathbb{E}\left[|e_t|\right]\,\mathrm{d}t\le J_R+\frac{2}{L}\sqrt{RM_R}. ( L R − 2 ) u H past ( L ) ≤ J R + L 2 ∫ 0 R E [ ∣ e t ∣ ] d t ≤ J R + L 2 R M R .
Multiplying by L / R L/R L / R to get
( 1 − 2 L R ) u H p a s t ( L ) ≤ L J R R + 2 M R R ≤ L J R R + 2 2 J R R . \left(1-\frac{2L}{R}\right)u^{\mathrm{past}}_H(L)\le L\frac{J_R}{R}+2\sqrt{\frac{M_R}{R}}\le L\frac{J_R}{R}+2\sqrt{\frac{2J_R}{R}}. ( 1 − R 2 L ) u H past ( L ) ≤ L R J R + 2 R M R ≤ L R J R + 2 R 2 J R .
Letting R → ∞ R\to\infty R → ∞ proves the upper bound. Finally,
κ H ≤ u H ( 1 ) ≤ 1 2 ( 2 H + 1 ) + m 1 ≤ 1 2 + m 1 , \kappa_H\le u_H(1)\le\frac{1}{2(2H+1)}+m_1\le\frac12+m_1, κ H ≤ u H ( 1 ) ≤ 2 ( 2 H + 1 ) 1 + m 1 ≤ 2 1 + m 1 ,
which gives the uniform constant C 0 C_0 C 0 .
Prediction states. For the one-sided fractional Brownian motion with its natural filtration, the observed past determines the prediction curve
G t ( r ) : = E [ B t + r H ∣ F t B H ] , r ≥ 0. G_t(r) := \mathbb{E}\left[B^H_{t+r}\mid\mathcal{F}^{B^H}_t\right], \qquad r \ge0. G t ( r ) := E [ B t + r H ∣ F t B H ] , r ≥ 0.
The conditional law of the future depends on the past only through ( t , G t ) (t,G_t) ( t , G t ) . Its Markov property and its role in continuation values are established in section G. A one-step tracking problem there identifies information relevant to a decision beyond the current scalar error.
Local homogenization
A filtered fractional tangent
In the rest of this section, the original probability space is standard up to completion and each F t \mathcal{F}_t F t is the completion of a countably generated sigma-field. These conditions hold on the canonical Gaussian path spaces used below.
Assumption 6.1 (filtered L 2 L^2 L 2 fractional tangent). There are H ∈ ( 0 , 1 ) H \in(0,1) H ∈ ( 0 , 1 ) and ρ ∈ C ( [ 0 , T ] ; ( 0 , ∞ ) ) \rho\in\mathcal{C}([0,T];(0,\infty)) ρ ∈ C ([ 0 , T ] ; ( 0 , ∞ )) such that the following three conditions hold. (FT1) Filtered coupling. For every compact J ⊂ ( 0 , T ) J \subset(0,T) J ⊂ ( 0 , T ) and every 0 < c − < c + < ∞ 0<c_-<c_+<\infty 0 < c − < c + < ∞ , there are h ∗ = h ∗ ( J , c − , c + ) > 0 h_* = h_*(J,c_-,c_+)>0 h ∗ = h ∗ ( J , c − , c + ) > 0 and a common extension ( Ω ‾ , F ‾ , P ‾ ) (\overline{\Omega},\overline{\mathcal{F}},\overline{\mathbb{P}}) ( Ω , F , P ) of the original probability space. If H ⊂ F \mathcal{H}\subset\mathcal{F} H ⊂ F , write H ‾ \overline{\mathcal{H}} H for its P ‾ \overline{\mathbb{P}} P -completion. On this extension, for every t ∈ J t\in J t ∈ J , every 0 < h < h ∗ 0<h<h_* 0 < h < h ∗ , and every c ∈ [ c − , c + ] c\in[c_-,c_+] c ∈ [ c − , c + ] , there is an auxiliary random variable U t , h , c U^{t,h,c} U t , h , c taking values in a standard Borel space, independent of the pullback of F T \mathcal{F}_T F T , and a two-sided fBm B t , h , c B^{t,h,c} B t , h , c with L ( B t , h , c ) = μ H p a s t \mathcal{L}(B^{t,h,c})=\mu_H^{\mathrm{past}} L ( B t , h , c ) = μ H past . For every u ∈ [ 0 , ( T − t ) / ( c h ) ] u\in[0,(T-t)/(ch)] u ∈ [ 0 , ( T − t ) / ( c h )] ,
F ‾ t + c h u ∨ σ ( U t , h , c ) = σ ( B v t , h , c : v ≤ u ) . (36) \overline{\mathcal{F}}_{t+chu}\vee\sigma(U^{t,h,c})=\sigma(B^{t,h,c}_v:v\le u). \tag*{(36)} F t + c h u ∨ σ ( U t , h , c ) = σ ( B v t , h , c : v ≤ u ) . (36) (FT2) Uniform L 2 L^2 L 2 tangent approximation. For every R < ∞ R<\infty R < ∞ ,
sup t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ R ∥ Θ t + c h u − Θ t ρ ( t ) ( c h ) H − B u t , h , c ∥ L 2 ( P ‾ ) → h → 0 0. (37) \sup_{\substack{t\in J,\ c\in[c_-,c_+]\\0\le u\le R}}
\left\|
\frac{\Theta_{t+chu}-\Theta_t}{\rho(t)(ch)^H}-B^{t,h,c}_u
\right\|_{L^2(\overline{\mathbb{P}})}
\xrightarrow[h\to0]{}0. \tag*{(37)} t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ R sup ρ ( t ) ( c h ) H Θ t + c h u − Θ t − B u t , h , c L 2 ( P ) h → 0 0. (37) (FT3) Global increment bound. For some C , h 0 > 0 C,h_0>0 C , h 0 > 0 ,
sup 0 < h < h 0 0 ≤ t ≤ T − h h − 2 H E [ ∣ Θ t + h − Θ t ∣ 2 ] ≤ C . (38) \sup_{\substack{0<h<h_0\\0\le t\le T-h}}
h^{-2H}\mathbb{E}\left[|\Theta_{t+h}-\Theta_t|^2\right]\le C. \tag*{(38)} 0 < h < h 0 0 ≤ t ≤ T − h sup h − 2 H E [ ∣ Θ t + h − Θ t ∣ 2 ] ≤ C . (38)
Assume from now on that a , b , ρ ∈ C ( [ 0 , T ] ; ( 0 , ∞ ) ) a,b,\rho\in\mathcal{C}([0,T];(0,\infty)) a , b , ρ ∈ C ([ 0 , T ] ; ( 0 , ∞ )) . Define α H : = 2 H / ( 1 + H ) ∈ ( 0 , 1 ) \alpha_H:=2H/(1+H)\in(0,1) α H := 2 H / ( 1 + H ) ∈ ( 0 , 1 ) and
g H : { [ 0 , T ] ⟶ ( 0 , ∞ ) , t ⟼ a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ρ ( t ) 2 / ( 1 + H ) . (39) g_H:
\begin{cases}
[0,T]\longrightarrow(0,\infty),\\
t\longmapsto a_t^{(1-H)/(1+H)}b_t^{2H/(1+H)}\rho(t)^{2/(1+H)}.
\end{cases} \tag*{(39)} g H : { [ 0 , T ] ⟶ ( 0 , ∞ ) , t ⟼ a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ρ ( t ) 2/ ( 1 + H ) . (39)
Theorem 6.2 (local homogenization). Suppose assumption 6.1 holds and a , b , ρ a,b,\rho a , b , ρ are continuous and bounded away from zero. Then
λ − α H V λ ( Θ ) → λ → 0 κ H ∫ 0 T g H ( t ) d t . (40) \lambda^{-\alpha_H}V_\lambda(\Theta)\xrightarrow[\lambda\to0]{}\kappa_H\int_0^T g_H(t)\,\mathrm{d}t. \tag*{(40)} λ − α H V λ ( Θ ) λ → 0 κ H ∫ 0 T g H ( t ) d t . (40)
The proof is given below, after the reset-value estimates.
Stability of reset values
Fix L > 0 L>0 L > 0 , write G = ( G t ) 0 ≤ t ≤ L \mathbf{G}=(\mathcal{G}_t)_{0\leq t\leq L} G = ( G t ) 0 ≤ t ≤ L , and let ( Ω , F , G , P ) (\Omega,\mathcal{F},\mathbf{G},\mathbb{P}) ( Ω , F , G , P ) be a complete filtered probability space satisfying the usual conditions. Let X ∈ L 0 ( Ω ; C ( [ 0 , L ] ; R ) ) X\in L^0(\Omega;C([0,L];\mathbb{R})) X ∈ L 0 ( Ω ; C ([ 0 , L ] ; R )) be G \mathbf{G} G -adapted, with X 0 = 0 X_0=0 X 0 = 0 and E [ ∫ 0 L X t 2 d t + ∣ X L ∣ ] < ∞ \mathbb{E}\left[\int_0^L X_t^2\,\mathrm{d}t+\lvert X_L\rvert\right]<\infty E [ ∫ 0 L X t 2 d t + ∣ X L ∣ ] < ∞ . Let A L ( G ) \mathcal{A}_L(\mathbf{G}) A L ( G ) consist of the G \mathbf{G} G -adapted processes ψ ∈ L 0 ( Ω ; D L ) \psi\in L^0(\Omega;\mathbb{D}_L) ψ ∈ L 0 ( Ω ; D L ) such that ψ 0 − = 0 \psi_{0-}=0 ψ 0 − = 0 , TV ( 0 , L ] ( ψ ) < ∞ \operatorname{TV}_{(0,L]}(\psi)<\infty TV ( 0 , L ] ( ψ ) < ∞ almost surely, and
E [ ∫ 0 L ψ t 2 d t + TV [ 0 , L ] a u g ( ψ ) + ∣ ψ L ∣ ] < ∞ . \mathbb{E}\left[\int_0^L\psi_t^2\,\mathrm{d}t+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)+\lvert\psi_L\rvert\right]<\infty. E [ ∫ 0 L ψ t 2 d t + TV [ 0 , L ] aug ( ψ ) + ∣ ψ L ∣ ] < ∞.
For these controls, define the pathwise reset cost by
C L X : A L ( G ) ⟶ L 1 ( Ω ; [ 0 , ∞ ) ) ψ ⟼ 1 2 ∫ 0 L ( ψ t − X t ) 2 d t + TV [ 0 , L ] a u g ( ψ ) + ∣ ψ L − X L ∣ \mathcal{C}_L^X:\quad
\begin{array}{rcl}
\mathcal{A}_L(\mathbf{G})&\longrightarrow&L^1(\Omega;[0,\infty))\\
\psi&\longmapsto&\frac{1}{2}\int_0^L(\psi_t-X_t)^2\,\mathrm{d}t+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)+\lvert\psi_L-X_L\rvert
\end{array} C L X : A L ( G ) ψ ⟶ ⟼ L 1 ( Ω ; [ 0 , ∞ )) 2 1 ∫ 0 L ( ψ t − X t ) 2 d t + TV [ 0 , L ] aug ( ψ ) + ∣ ψ L − X L ∣
Its expectation defines
J L X : A L ( G ) ⟶ [ 0 , ∞ ) ψ ⟼ E [ C L X ( ψ ) ] J_L^X:\quad
\begin{array}{rcl}
\mathcal{A}_L(\mathbf{G})&\longrightarrow&[0,\infty)\\
\psi&\longmapsto&\mathbb{E}[\mathcal{C}_L^X(\psi)]
\end{array} J L X : A L ( G ) ψ ⟶ ⟼ [ 0 , ∞ ) E [ C L X ( ψ )]
so the reset value is
U L ( X ; G ) : = inf ψ ∈ A L ( G ) J L X ( ψ ) . (41) \mathcal{U}_L(X;\mathbf{G}):=\inf_{\psi\in\mathcal{A}_L(\mathbf{G})}J_L^X(\psi). \tag*{(41)} U L ( X ; G ) := ψ ∈ A L ( G ) inf J L X ( ψ ) . (41)
When the filtration is clear, write U L ( X ) : = U L ( X ; G ) \mathcal{U}_L(X):=\mathcal{U}_L(X;\mathbf{G}) U L ( X ) := U L ( X ; G ) .
Lemma 6.3 (stability under a common filtration). Let X , Y X,Y X , Y start at zero and be adapted to the same filtration G \mathbf{G} G , with finite reset values. Put A L : = E [ ∫ 0 L ∣ X t − Y t ∣ 2 d t ] A_L:=\mathbb{E}\left[\int_0^L\lvert X_t-Y_t\rvert^2\,\mathrm{d}t\right] A L := E [ ∫ 0 L ∣ X t − Y t ∣ 2 d t ] and B L : = E [ ∣ X L − Y L ∣ ] B_L:=\mathbb{E}[\lvert X_L-Y_L\rvert] B L := E [∣ X L − Y L ∣] . Then, for every δ > 0 \delta>0 δ > 0 ,
U L ( X ; G ) ≤ ( 1 + δ ) U L ( Y ; G ) + 1 2 ( 1 + δ − 1 ) A L + B L . (42) \mathcal{U}_L(X;\mathbf{G})\leq(1+\delta)\mathcal{U}_L(Y;\mathbf{G})+\frac{1}{2}(1+\delta^{-1})A_L+B_L. \tag*{(42)} U L ( X ; G ) ≤ ( 1 + δ ) U L ( Y ; G ) + 2 1 ( 1 + δ − 1 ) A L + B L . (42)
More precisely , writing U X = U L ( X ; G ) U_X=\mathcal{U}_L(X;\mathbf{G}) U X = U L ( X ; G ) and U Y = U L ( Y ; G ) U_Y=\mathcal{U}_L(Y;\mathbf{G}) U Y = U L ( Y ; G ) ,
∣ U X − U Y ∣ ≤ B L + A L 2 + 2 A L min ( U X , U Y ) . (43) \lvert U_X-U_Y\rvert\leq B_L+\frac{A_L}{2}+\sqrt{2A_L}\min(U_X,U_Y). \tag*{(43)} ∣ U X − U Y ∣ ≤ B L + 2 A L + 2 A L min ( U X , U Y ) . (43)
Therefore , if both values are O ( L ) O(L) O ( L ) and
A L L → L → ∞ 0 , B L L → L → ∞ 0 , (44) \frac{A_L}{L}\xrightarrow[L\to\infty]{}0,\qquad\frac{B_L}{L}\xrightarrow[L\to\infty]{}0, \tag*{(44)} L A L L → ∞ 0 , L B L L → ∞ 0 , (44)
then ∣ U X − U Y ∣ = o ( L ) \lvert U_X-U_Y\rvert=o(L) ∣ U X − U Y ∣ = o ( L ) .
Proof. See section H.1 .
Lemma 6.4 (removing conditionally independent auxiliary variables). Let E E E and S S S be standard Borel spaces. Assume that the probability space is standard up to completion and that each G t \mathcal{G}_t G t is the completion of a countably generated sigma-field. Let Y ∈ L 0 ( Ω , G 0 , P ; E ) Y\in L^0(\Omega,\mathcal{G}_0,\mathbb{P};E) Y ∈ L 0 ( Ω , G 0 , P ; E ) and U ∈ L 0 ( Ω , F , P ; S ) U\in L^0(\Omega,\mathcal{F},\mathbb{P};S) U ∈ L 0 ( Ω , F , P ; S ) satisfy
U ⊥ ⊥ G L ∣ Y . (45) U\perp\!\!\!\perp\mathcal{G}_L\mid Y. \tag*{(45)} U ⊥ ⊥ G L ∣ Y . (45)
Write G t U : = G t ∨ σ ( U ) P \mathcal{G}_t^U:=\mathcal{G}_t\vee\sigma(U)^{\mathbb{P}} G t U := G t ∨ σ ( U ) P and let G U : = ( G t U ) 0 ≤ t ≤ L \mathbf{G}^U:=(\mathcal{G}_t^U)_{0\leq t\leq L} G U := ( G t U ) 0 ≤ t ≤ L be the enlarged filtration. Then
U L ( X ; G U ) = U L ( X ; G ) . (46) \mathcal{U}_L(X;\mathbf{G}^U)=\mathcal{U}_L(X;\mathbf{G}). \tag*{(46)} U L ( X ; G U ) = U L ( X ; G ) . (46)
More precisely ,
ess inf ψ ∈ A L ( G U ) E [ C L X ( ψ ) ∣ Y ] = ess inf ψ ∈ A L ( G ) E [ C L X ( ψ ) ∣ Y ] P -a.s. (47) \operatorname*{ess\,inf}_{\psi\in\mathcal{A}_L(\mathbf{G}^U)}
\mathbb{E}[\mathcal{C}_L^X(\psi)\mid Y]
=
\operatorname*{ess\,inf}_{\psi\in\mathcal{A}_L(\mathbf{G})}
\mathbb{E}[\mathcal{C}_L^X(\psi)\mid Y]
\quad\mathbb{P}\text{-a.s.} \tag*{(47)} ψ ∈ A L ( G U ) ess inf E [ C L X ( ψ ) ∣ Y ] = ψ ∈ A L ( G ) ess inf E [ C L X ( ψ ) ∣ Y ] P -a.s. (47)
The conditional values at the initial time also agree:
ess inf ψ ∈ A L ( G U ) E [ C L X ( ψ ) ∣ G 0 U ] = ess inf ψ ∈ A L ( G ) E [ C L X ( ψ ) ∣ G 0 ] . (48) \operatorname*{ess\,inf}_{\psi\in\mathcal{A}_L(\mathbf{G}^U)}
\mathbb{E}[\mathcal{C}_L^X(\psi)\mid\mathcal{G}_0^U]
=
\operatorname*{ess\,inf}_{\psi\in\mathcal{A}_L(\mathbf{G})}
\mathbb{E}[\mathcal{C}_L^X(\psi)\mid\mathcal{G}_0]. \tag*{(48)} ψ ∈ A L ( G U ) ess inf E [ C L X ( ψ ) ∣ G 0 U ] = ψ ∈ A L ( G ) ess inf E [ C L X ( ψ ) ∣ G 0 ] . (48)
The same statements hold for the deterministic weighted tracking costs used in this paper and for a pre-initial position measurable with respect to G 0 \mathcal{G}_0 G 0 , if some admissible control has finite cost.
Proof. See section B .
Uniform growing-cell approximation
Lemma 6.5 (uniform growing-cell approximation). Under assumption 6.1, fix a compact J ⊂ ( 0 , T ) J\subset(0,T) J ⊂ ( 0 , T ) and 0 < c − < c + < ∞ 0<c_-<c_+<\infty 0 < c − < c + < ∞ . For admissible ( t , h , c ) (t,h,c) ( t , h , c ) , define the local target
X t , h , c : { [ 0 , ( T − t ) / ( c h ) ] ⟶ L 2 ( Ω ) , u ⟼ Θ t + c h u − Θ t ρ ( t ) ( c h ) H , X^{t,h,c}:\begin{cases}
[0,(T-t)/(ch)]\longrightarrow L^2(\Omega),\\
u\longmapsto\dfrac{\Theta_{t+chu}-\Theta_t}{\rho(t)(ch)^H},
\end{cases} X t , h , c : ⎩ ⎨ ⎧ [ 0 , ( T − t ) / ( c h )] ⟶ L 2 ( Ω ) , u ⟼ ρ ( t ) ( c h ) H Θ t + c h u − Θ t , and define the local filtration G ‾ t , h , c = ( G ‾ u t , h , c ) 0 ≤ u ≤ ( T − t ) / ( c h ) \overline{\mathcal{G}}^{t,h,c}=(\overline{\mathcal{G}}^{t,h,c}_u)_{0\leq u\leq(T-t)/(ch)} G t , h , c = ( G u t , h , c ) 0 ≤ u ≤ ( T − t ) / ( c h ) by
G ‾ u t , h , c : = F t + c h u ∨ σ ( U t t , h , c ) . \overline{\mathcal{G}}^{t,h,c}_u:=\mathcal{F}_{t+chu}\vee\sigma(U^{t,h,c}_t). G u t , h , c := F t + c h u ∨ σ ( U t t , h , c ) . By lemma 6.4, the value of the original target is the same with or without this enlargement. There is L h ↑ ∞ L_h\uparrow\infty L h ↑ ∞ such that
h L h → h → 0 0 hL_h\xrightarrow[h\to0]{}0 h L h h → 0 0 and, uniformly in t ∈ J t\in J t ∈ J and c ∈ [ c − , c + ] c\in[c_-,c_+] c ∈ [ c − , c + ] ,
1 L h L ( ( Θ t + c h u − Θ t ρ ( t ) ( c h ) H ) 0 ≤ u ≤ L h ; ( F t + c h u ) u ≤ L h ) → h → 0 κ H . (49) \frac{1}{L_h}\mathcal{L}\left(\left(\frac{\Theta_{t+chu}-\Theta_t}{\rho(t)(ch)^H}\right)_{0\leq u\leq L_h};(\mathcal{F}_{t+chu})_{u\leq L_h}\right)\xrightarrow[h\to0]{}\kappa_H. \tag*{(49)} L h 1 L ( ( ρ ( t ) ( c h ) H Θ t + c h u − Θ t ) 0 ≤ u ≤ L h ; ( F t + c h u ) u ≤ L h ) h → 0 κ H . (49)
Proof. See section H.2.
Proof of the local homogenization theorem
The block reset decomposition, separator points, and short-interval control used below are established in section H.3.
Proof of theorem 6.2. We first choose the local scales and establish a uniform block comparison. For f ∈ C ( [ 0 , T ] ; R ) f\in\mathcal{C}([0,T];\mathbb{R}) f ∈ C ([ 0 , T ] ; R ) , write its modulus of continuity as
ω f : { [ 0 , ∞ ) ⟶ [ 0 , ∞ ) , r ⟼ sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ r ∣ f ( t ) − f ( s ) ∣ . \omega_f:\begin{cases}
[0,\infty)\longrightarrow[0,\infty),\\
r\longmapsto\sup_{\substack{s,t\in[0,T]\\ |t-s|\leq r}}|f(t)-f(s)|.
\end{cases} ω f : ⎩ ⎨ ⎧ [ 0 , ∞ ) ⟶ [ 0 , ∞ ) , r ⟼ sup s , t ∈ [ 0 , T ] ∣ t − s ∣ ≤ r ∣ f ( t ) − f ( s ) ∣.
Put h λ : = λ 1 / ( 1 + H ) h_\lambda:=\lambda^{1/(1+H)} h λ := λ 1/ ( 1 + H ) , so λ = h λ 1 + H \lambda=h_\lambda^{1+H} λ = h λ 1 + H and λ α H = h λ 2 H \lambda^{\alpha H}=h_\lambda^{2H} λ α H = h λ 2 H , and define
ℓ λ : { [ 0 , T ] ⟶ ( 0 , ∞ ) , t ⟼ ( λ b t a t ρ ( t ) ) 1 / ( 1 + H ) . (50) \ell_\lambda:\begin{cases}
[0,T]\longrightarrow(0,\infty),\\
t\longmapsto\left(\dfrac{\lambda b_t}{a_t\rho(t)}\right)^{1/(1+H)}.
\end{cases} \tag*{(50)} ℓ λ : ⎩ ⎨ ⎧ [ 0 , T ] ⟶ ( 0 , ∞ ) , t ⟼ ( a t ρ ( t ) λ b t ) 1/ ( 1 + H ) . (50)
Figure 2 illustrates this time and amplitude normalization. Positivity and continuity give constants 0 < c − < c + < ∞ 0<c_-<c_+<\infty 0 < c − < c + < ∞ with
Figure 2. Local rescaling of a tracking block. The physical interval I i = [ t i , t i + ℓ i R ] I_i=[t_i,t_i+\ell_iR] I i = [ t i , t i + ℓ i R ] becomes [ 0 , R ] [0,R] [ 0 , R ] , and target increments are normalized by ρ i ℓ i H \rho_i\ell_i^H ρ i ℓ i H . The lower arrow represents the filtered L 2 L^2 L 2 coupling with the two-sided fBm tangent.
ℓ λ ( t ) h λ = ( b t a t ρ ( t ) ) 1 / ( 1 + H ) ∈ [ c − , c + ] . \frac{\ell_\lambda(t)}{h_\lambda}=\left(\frac{b_t}{a_t\rho(t)}\right)^{1/(1+H)}\in[c_-,c_+]. h λ ℓ λ ( t ) = ( a t ρ ( t ) b t ) 1/ ( 1 + H ) ∈ [ c − , c + ] .
Fix η ∈ ( 0 , T / 2 ) \eta\in(0,T/2) η ∈ ( 0 , T /2 ) and apply lemma 6.5 on J η : = [ η / 2 , T − η / 2 ] J_\eta:= [\eta/2,T-\eta/2] J η := [ η /2 , T − η /2 ] . Write L λ : = L h λ L_\lambda:= L_{h_\lambda} L λ := L h λ and Δ λ : = h λ L λ \Delta_\lambda:= h_\lambda L_\lambda Δ λ := h λ L λ . Then
L λ → λ → 0 ∞ , Δ λ → λ → 0 0. L_\lambda\xrightarrow[\lambda\to0]{} \infty,\qquad\Delta_\lambda\xrightarrow[\lambda\to0]{} 0. L λ λ → 0 ∞ , Δ λ λ → 0 0.
The extension to comparable horizons is proved in section H.2. Uniformly for starts in J η J_\eta J η and horizons R = q L λ + O ( 1 ) R=qL_\lambda+O(1) R = q L λ + O ( 1 ) , with q q q bounded above and away from zero,
R R ( ( Θ s + ℓ λ ( s ) u − Θ s ρ ( s ) ℓ λ ( s ) H ) 0 ≤ u ≤ R ; ( F s + ℓ λ ( s ) u ) 0 ≤ u ≤ R ) = ( κ H + o ( 1 ) ) R . \mathcal{R}_R\left(\left(\frac{\Theta_{s+\ell_\lambda(s)u}-\Theta_s}{\rho(s)\ell_\lambda(s)H}\right)_{0\leq u\leq R};(\mathcal{F}_{s+\ell_\lambda(s)u})_{0\leq u\leq R}\right)=(\kappa_H+o(1))R. R R ( ( ρ ( s ) ℓ λ ( s ) H Θ s + ℓ λ ( s ) u − Θ s ) 0 ≤ u ≤ R ; ( F s + ℓ λ ( s ) u ) 0 ≤ u ≤ R ) = ( κ H + o ( 1 )) R .
To bound the effect of holding the coefficients constant, put C ∗ : = 2 ( 1 ∨ c + ) C_*:=2(1\vee c_+) C ∗ := 2 ( 1 ∨ c + ) and set
ω λ : = max f ∈ { a , b , ρ } ω f ( C ∗ Δ λ ) min [ 0 , T ] f → λ → 0 0. \omega_\lambda:=\max_{f\in\{a,b,\rho\}}\frac{\omega_f(C_*\Delta_\lambda)}{\min_{[0,T]}f}\xrightarrow[\lambda\to0]{}0. ω λ := f ∈ { a , b , ρ } max min [ 0 , T ] f ω f ( C ∗ Δ λ ) λ → 0 0.
Holding the coefficients constant on a block of length at most C ∗ Δ λ C_*\Delta_\lambda C ∗ Δ λ changes every nonnegative cost term by a relative factor between 1 − ω λ 1-\omega_\lambda 1 − ω λ and 1 + ω λ 1+\omega_\lambda 1 + ω λ . To obtain the upper bound, start at s 0 : = η s_0:=\eta s 0 := η and take consecutive blocks I i : = [ s i , s i + 1 ] ⊂ [ η , T − η ] I_i:=[s_i,s_{i+1}]\subset[\eta,T-\eta] I i := [ s i , s i + 1 ] ⊂ [ η , T − η ] of length ℓ i L λ \ell_iL_\lambda ℓ i L λ , where ℓ i : = ℓ λ ( s i ) \ell_i:=\ell_\lambda(s_i) ℓ i := ℓ λ ( s i ) . Let M λ M_\lambda M λ be the number of these blocks, indexed by i = 0 , … , M λ − 1 i=0,\ldots,M_\lambda-1 i = 0 , … , M λ − 1 . For each block, write a i : = a s i a_i:=a_{s_i} a i := a s i , b i : = b s i b_i:=b_{s_i} b i := b s i and ρ i : = ρ ( s i ) \rho_i:=\rho(s_i) ρ i := ρ ( s i ) . There are O ( Δ λ − 1 ) O(\Delta_\lambda^{-1}) O ( Δ λ − 1 ) blocks, and the interior remainder has length O ( Δ λ ) O(\Delta_\lambda) O ( Δ λ ) . With t = s i + ℓ i u t=s_i+\ell_i u t = s i + ℓ i u and ϕ t = Θ s i + ρ i ℓ i ψ u \phi_t=\Theta_{s_i}+\rho_i\ell_i\psi_u ϕ t = Θ s i + ρ i ℓ i ψ u , the reset cost J i , λ f i x J^{\mathrm{fix}}_{i,\lambda} J i , λ fix with coefficients held constant satisfies
inf ψ J i , λ f i x ( ψ ) = a i ρ i 2 ℓ i 2 H + 1 U L λ ( ( Θ s i + ℓ i u − Θ s i ρ i ℓ i H ) 0 ≤ u ≤ L λ ) . (51) \inf_\psi J^{\mathrm{fix}}_{i,\lambda}(\psi)=a_i\rho_i^2\ell_i^{2H+1}\mathcal{U}_{L_\lambda}\left(\left(\frac{\Theta_{s_i+\ell_i u}-\Theta_{s_i}}{\rho_i\ell_i^H}\right)_{0\leq u\leq L_\lambda}\right). \tag*{(51)} ψ inf J i , λ fix ( ψ ) = a i ρ i 2 ℓ i 2 H + 1 U L λ ( ( ρ i ℓ i H Θ s i + ℓ i u − Θ s i ) 0 ≤ u ≤ L λ ) . (51)
Indeed, ℓ i H + 1 = λ b i / ( a i ρ i ) \ell_i^{H+1}=\lambda b_i/(a_i\rho_i) ℓ i H + 1 = λ b i / ( a i ρ i ) , so the running and variation/reset coefficients agree:
a i ρ i 2 ℓ i 2 H + 1 = λ b i ρ i ℓ i H . a_i\rho_i^2\ell_i^{2H+1}=\lambda b_i\rho_i\ell_i^H. a i ρ i 2 ℓ i 2 H + 1 = λ b i ρ i ℓ i H .
Moreover,
a i ρ i 2 ℓ i 2 H = λ α H a i 1 − α H b i α H ρ i 2 − α H = λ α H g H ( s i ) . (52) a_i\rho_i^2\ell_i^{2H}=\lambda^{\alpha_H}a_i^{1-\alpha_H}b_i^{\alpha_H}\rho_i^{2-\alpha_H}=\lambda^{\alpha_H}g_H(s_i). \tag*{(52)} a i ρ i 2 ℓ i 2 H = λ α H a i 1 − α H b i α H ρ i 2 − α H = λ α H g H ( s i ) . (52)
Since ∣ I i ∣ = ℓ i L λ |I_i|=\ell_iL_\lambda ∣ I i ∣ = ℓ i L λ , the uniform cell comparison gives
inf ψ J i , λ f i x ( ψ ) = ( κ H + o ( 1 ) ) λ α H g H ( s i ) ∣ I i ∣ . \inf_\psi J^{\mathrm{fix}}_{i,\lambda}(\psi)=(\kappa_H+o(1))\lambda^{\alpha_H}g_H(s_i)|I_i|. ψ inf J i , λ fix ( ψ ) = ( κ H + o ( 1 )) λ α H g H ( s i ) ∣ I i ∣.
Choose original-filtration controls within δ λ L λ \delta_\lambda L_\lambda δ λ L λ of the normalized infima, where δ λ → λ → 0 0 \delta_\lambda\xrightarrow[\lambda\to0]{}0 δ λ λ → 0 0 . Lemma 6.4 removes the auxiliary enlargement. Their total approximation error is
∑ i = 0 M λ − 1 a i ρ i 2 ℓ i 2 H + 1 δ λ L λ = δ λ λ α H ∑ i = 0 M λ − 1 g H ( s i ) ∣ I i ∣ = O ( δ λ λ α H ) = o ( λ α H ) . \sum_{i=0}^{M_\lambda-1}a_i\rho_i^2\ell_i^{2H+1}\delta_\lambda L_\lambda
=\delta_\lambda\lambda^{\alpha_H}\sum_{i=0}^{M_\lambda-1}g_H(s_i)|I_i|
=O(\delta_\lambda\lambda^{\alpha_H})=o(\lambda^{\alpha_H}). i = 0 ∑ M λ − 1 a i ρ i 2 ℓ i 2 H + 1 δ λ L λ = δ λ λ α H i = 0 ∑ M λ − 1 g H ( s i ) ∣ I i ∣ = O ( δ λ λ α H ) = o ( λ α H ) .
The cost with the original coefficients is at most 1 + ω λ 1+\omega_\lambda 1 + ω λ times the cost with coefficients held constant. Terminal resets and initial atoms cover the cost of all jumps between blocks (lemma C.3). Therefore the interior cost is at most
( κ H + o ( 1 ) ) λ α H ∑ i = 0 M λ − 1 g H ( s i ) ∣ I i ∣ . (\kappa_H+o(1))\lambda^{\alpha_H}\sum_{i=0}^{M_\lambda-1}g_H(s_i)|I_i|. ( κ H + o ( 1 )) λ α H i = 0 ∑ M λ − 1 g H ( s i ) ∣ I i ∣.
The Riemann sums satisfy
∑ i = 0 M λ − 1 g H ( s i ) ∣ I i ∣ → λ → 0 ∫ η T − η g H ( t ) d t . \sum_{i=0}^{M_\lambda-1}g_H(s_i)|I_i|\xrightarrow[\lambda\to0]{}\int_\eta^{T-\eta}g_H(t)\,\mathrm{d}t. i = 0 ∑ M λ − 1 g H ( s i ) ∣ I i ∣ λ → 0 ∫ η T − η g H ( t ) d t .
Use lemma H.3 on [ 0 , η ] [0,\eta] [ 0 , η ] , [ T − η , T ] [T-\eta,T] [ T − η , T ] , and the interior remainder. Their combined cost is at most C λ α H ( 2 η + O ( Δ λ ) + 3 h λ ) C\lambda^{\alpha_H}(2\eta+O(\Delta_\lambda)+3h_\lambda) C λ α H ( 2 η + O ( Δ λ ) + 3 h λ ) , whence
lim sup λ → 0 λ − α H V λ ( Θ ) ≤ κ H ∫ η T − η g H ( t ) d t + C η . \limsup_{\lambda\to0}\lambda^{-\alpha_H}V_\lambda(\Theta)\leq\kappa_H\int_\eta^{T-\eta}g_H(t)\,\mathrm{d}t+C\eta. λ → 0 lim sup λ − α H V λ ( Θ ) ≤ κ H ∫ η T − η g H ( t ) d t + C η .
For the lower bound, we first control the extra cost of separating blocks. Take ϕ λ \phi^\lambda ϕ λ with cost V λ ( Θ ) + o ( λ α H ) V_\lambda(\Theta)+o(\lambda^{\alpha_H}) V λ ( Θ ) + o ( λ α H ) . The upper bound and a ≥ a − > 0 a\geq a_->0 a ≥ a − > 0 give
E [ ∫ 0 T ∣ ϕ t λ − Θ t ∣ 2 d t ] ≤ 2 a − J λ ( ϕ λ ; Θ ) ≤ C λ α H = C h λ 2 H . (53) \mathbb{E}\left[\int_0^T|\phi_t^\lambda-\Theta_t|^2\,\mathrm{d}t\right]\leq\frac{2}{a_-}J_\lambda(\phi^\lambda;\Theta)\leq C\lambda^{\alpha_H}=Ch_\lambda^{2H}. \tag*{(53)} E [ ∫ 0 T ∣ ϕ t λ − Θ t ∣ 2 d t ] ≤ a − 2 J λ ( ϕ λ ; Θ ) ≤ C λ α H = C h λ 2 H . (53)
For sufficiently small λ \lambda λ , set K λ : = ⌊ ( T − 2 η − h λ ) / Δ λ ⌋ K_\lambda:= \lfloor(T - 2\eta- h_\lambda)/\Delta_\lambda\rfloor K λ := ⌊( T − 2 η − h λ ) / Δ λ ⌋ and use grid points z j : = η + j Δ λ z_j := \eta+ j\Delta_\lambda z j := η + j Δ λ , j = 0 , … , K λ j = 0, \ldots, K_\lambda j = 0 , … , K λ , so that J j : = [ z j , z j + h λ ] ⊂ [ η , T − η ] J_j := [z_j,z_j+h_\lambda] \subset[\eta,T-\eta] J j := [ z j , z j + h λ ] ⊂ [ η , T − η ] . For small λ \lambda λ , these intervals are disjoint. Their number is O ( Δ λ − 1 ) O(\Delta_\lambda^{-1}) O ( Δ λ − 1 ) . By lemma H.2, choose deterministic r j ∈ J j r_j \in J_j r j ∈ J j such that
∑ j = 0 K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] ≤ ( K λ + 1 h λ E [ ∫ 0 T ∣ ϕ t λ − Θ t ∣ 2 d t ] ) 1 / 2 ≤ C ( h λ 2 H h λ 2 Λ λ ) 1 / 2 = C h λ H − 1 Λ λ . (54) \begin{aligned}
\sum_{j=0}^{K_\lambda} \mathbb{E}\left[|\phi^\lambda_{r_j}-\Theta_{r_j}|\right] \le\left(\frac{K_\lambda+1}{h_\lambda}\mathbb{E}\left[\int_0^T |\phi^\lambda_t-\Theta_t|^2\,\mathrm{d}t\right]\right)^{1/2} \\
&\le C\left(\frac{h_\lambda^{2H}}{h_\lambda^2\Lambda_\lambda}\right)^{1/2}=C\frac{h_\lambda^{H-1}}{\sqrt{\Lambda_\lambda}}. \tag*{(54)}
\end{aligned} j = 0 ∑ K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] ≤ ( h λ K λ + 1 E [ ∫ 0 T ∣ ϕ t λ − Θ t ∣ 2 d t ] ) 1/2 ≤ C ( h λ 2 Λ λ h λ 2 H ) 1/2 = C Λ λ h λ H − 1 . (54)
Let I i : = [ r i − 1 , r i ] I_i := [r_{i-1},r_i] I i := [ r i − 1 , r i ] , i = 1 , … , K λ i=1,\ldots,K_\lambda i = 1 , … , K λ . Add a virtual initial adjustment from Θ r i − 1 \Theta_{r_{i-1}} Θ r i − 1 and a terminal reset to Θ r i \Theta_{r_i} Θ r i . By (117) , their total excess cost is at most
2 λ b + ∑ j = 0 K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] = O ( h λ 1 + H h λ H − 1 Λ λ ) = O ( h λ 2 H Λ λ ) = o ( λ α H ) . (55) 2\lambda b_+ \sum_{j=0}^{K_\lambda}\mathbb{E}\left[|\phi^\lambda_{r_j}-\Theta_{r_j}|\right]
=O\left(\frac{h_\lambda^{1+H}h_\lambda^{H-1}}{\sqrt{\Lambda_\lambda}}\right)
=O\left(\frac{h_\lambda^{2H}}{\sqrt{\Lambda_\lambda}}\right)
=o(\lambda^{\alpha_H}). \tag*{(55)} 2 λ b + j = 0 ∑ K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] = O ( Λ λ h λ 1 + H h λ H − 1 ) = O ( Λ λ h λ 2 H ) = o ( λ α H ) . (55)
It remains to compare the values of the resulting blocks. Since r j ∈ [ z j , z j + h λ ] r_j\in[z_j,z_j+h_\lambda] r j ∈ [ z j , z j + h λ ] and z i − z i − 1 = Δ λ z_i-z_{i-1}=\Delta_\lambda z i − z i − 1 = Δ λ ,
∣ I i ∣ = Δ λ + O ( h λ ) = h λ Λ λ + O ( h λ ) . |I_i|=\Delta_\lambda+O(h_\lambda)=h_\lambda\Lambda_\lambda+O(h_\lambda). ∣ I i ∣ = Δ λ + O ( h λ ) = h λ Λ λ + O ( h λ ) .
For each block, set t i : = r i − 1 t_i:=r_{i-1} t i := r i − 1 and ℓ i − : = ℓ λ ( t i ) \ell_i^-:=\ell_\lambda(t_i) ℓ i − := ℓ λ ( t i ) . Its normalized horizon is
R i , λ : = ∣ I i ∣ ℓ i − = h λ ℓ i − Λ λ + O ( 1 ) , R_{i,\lambda}:=\frac{|I_i|}{\ell_i^-}=\frac{h_\lambda}{\ell_i^-}\Lambda_\lambda+O(1), R i , λ := ℓ i − ∣ I i ∣ = ℓ i − h λ Λ λ + O ( 1 ) ,
with h λ / ℓ i − ∈ [ 1 / c + , 1 / c − ] h_\lambda/\ell_i^- \in[1/c_+,1/c_-] h λ / ℓ i − ∈ [ 1/ c + , 1/ c − ] . Let J i λ J_i^\lambda J i λ be the expected weighted reset cost when restricting ϕ λ \phi^\lambda ϕ λ to I i I_i I i . Holding the coefficients constant and applying the cell estimate for comparable horizons give, uniformly in i i i ,
J i λ ≥ ( 1 − ω λ ) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H + 1 × U R i , λ ( ( Θ t i + ℓ i − u − Θ t i ρ ( t i ) ( ℓ i − ) H ) 0 ≤ u ≤ R i , λ ) ≥ ( κ H − o ( 1 ) ) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H + 1 R i , λ = ( κ H − o ( 1 ) ) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H ∣ I i ∣ = ( κ H − o ( 1 ) ) λ α H g H ( t i ) ∣ I i ∣ . \begin{aligned}
J_i^\lambda\ge(1-\omega_\lambda)a_{t_i}\rho(t_i)^2(\ell_i^-)^{2H+1}\times\mathcal{U}_{R_{i,\lambda}}\left(\left(\frac{\Theta_{t_i+\ell_i^-u}-\Theta_{t_i}}{\rho(t_i)(\ell_i^-)^H}\right)_{0\le u\le R_{i,\lambda}}\right) \\
&\ge(\kappa_H-o(1))a_{t_i}\rho(t_i)^2(\ell_i^-)^{2H+1}R_{i,\lambda} \\
&=(\kappa_H-o(1))a_{t_i}\rho(t_i)^2(\ell_i^-)^{2H}|I_i| \\
&=(\kappa_H-o(1))\lambda^{\alpha_H}g_H(t_i)|I_i|.
\end{aligned} J i λ ≥ ( 1 − ω λ ) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H + 1 × U R i , λ ( ( ρ ( t i ) ( ℓ i − ) H Θ t i + ℓ i − u − Θ t i ) 0 ≤ u ≤ R i , λ ) ≥ ( κ H − o ( 1 )) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H + 1 R i , λ = ( κ H − o ( 1 )) a t i ρ ( t i ) 2 ( ℓ i − ) 2 H ∣ I i ∣ = ( κ H − o ( 1 )) λ α H g H ( t i ) ∣ I i ∣.
All discarded original costs are nonnegative. Hence
J λ ( ϕ λ ; Θ ) ≥ ∑ i = 1 K λ J i λ − 2 λ b + + ∑ j = 0 K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] . J_\lambda(\phi^\lambda;\Theta)\ge\sum_{i=1}^{K_\lambda}J_i^\lambda-2\lambda b_++\sum_{j=0}^{K_\lambda}\mathbb{E}\left[|\phi^\lambda_{r_j}-\Theta_{r_j}|\right]. J λ ( ϕ λ ; Θ ) ≥ i = 1 ∑ K λ J i λ − 2 λ b + + j = 0 ∑ K λ E [ ∣ ϕ r j λ − Θ r j ∣ ] .
The selected blocks leave O ( Δ λ + h λ ) O(\Delta_\lambda+h_\lambda) O ( Δ λ + h λ ) of [ η , T − η ] [\eta,T-\eta] [ η , T − η ] uncovered, and their mesh tends to zero. Therefore,
∑ i = 1 K λ g H ( t i ) ∣ I i ∣ → λ → 0 ∫ η T − η g H ( t ) d t . \sum_{i=1}^{K_\lambda}g_H(t_i)|I_i|\xrightarrow[\lambda\to0]{}\int_\eta^{T-\eta}g_H(t)\,\mathrm{d}t. i = 1 ∑ K λ g H ( t i ) ∣ I i ∣ λ → 0 ∫ η T − η g H ( t ) d t .
Together with (55) , this yields
lim inf λ → 0 λ − α H V λ ( Θ ) ≥ κ H ∫ η T − η g H ( t ) d t . \liminf_{\lambda\to0}\lambda^{-\alpha_H}V_\lambda(\Theta)\ge\kappa_H\int_\eta^{T-\eta}g_H(t)\,\mathrm{d}t. λ → 0 lim inf λ − α H V λ ( Θ ) ≥ κ H ∫ η T − η g H ( t ) d t .
Let η → 0 \eta\to0 η → 0 in both bounds and use continuity of g H g_H g H .
The cost split of near-optimal controls
Set p H = ( 1 − H ) / ( 1 + H ) p_H=(1-H)/(1+H) p H = ( 1 − H ) / ( 1 + H ) , so that p H + α H = 1 p_H+\alpha_H=1 p H + α H = 1 .
Theorem 6.6 (local split between tracking and variation costs). Under the assumptions of theorem 6.2, let ϕ λ \phi^\lambda ϕ λ satisfy J λ ( ϕ λ ; Θ ) = V λ ( Θ ) + o ( λ α H ) J_\lambda(\phi^\lambda;\Theta)=V_\lambda(\Theta)+o(\lambda^{\alpha_H}) J λ ( ϕ λ ; Θ ) = V λ ( Θ ) + o ( λ α H ) . Define the finite tracking measure on [ 0 , T ] [0,T] [ 0 , T ] by
μ λ ( d t ) : = λ − α H a t 2 E [ ( ϕ t λ − Θ t ) 2 ] d t , (56) \mu_\lambda(\mathrm{d}t):=\lambda^{-\alpha_H}\frac{a_t}{2}\mathbb{E}\left[(\phi^\lambda_t-\Theta_t)^2\right]\mathrm{d}t, \tag*{(56)} μ λ ( d t ) := λ − α H 2 a t E [ ( ϕ t λ − Θ t ) 2 ] d t , (56)
and define the variation measure, for A ∈ B ( [ 0 , T ] ) A \in\mathcal{B}([0,T]) A ∈ B ([ 0 , T ]) , by
ν λ ( A ) : = λ 1 − α H E [ ∫ A b t ∣ d ϕ λ ∣ t ] . (57) \nu_\lambda(A) := \lambda^{1-\alpha_H}\mathbb{E}\left[\int_A b_t \lvert d\phi^\lambda\rvert_t\right]. \tag*{(57)} ν λ ( A ) := λ 1 − α H E [ ∫ A b t ∣ d ϕ λ ∣ t ] . (57)
Then, weakly as finite measures,
μ λ → λ → 0 w p H κ H g H ( t ) d t , and ν λ → λ → 0 w α H κ H g H ( t ) d t . (58) \mu_\lambda\xrightarrow[\lambda\to0]{\mathrm{w}} p_H\kappa_H g_H(t)\,dt, \qquad\text{and} \qquad\nu_\lambda\xrightarrow[\lambda\to0]{\mathrm{w}} \alpha_H\kappa_H g_H(t)\,dt. \tag*{(58)} μ λ w λ → 0 p H κ H g H ( t ) d t , and ν λ w λ → 0 α H κ H g H ( t ) d t . (58)
In particular, the leading fractions of the total cost are p H p_H p H for tracking and α H \alpha_H α H for variation.
Proof. Fix f ∈ C ( [ 0 , T ] ; R ) f \in\mathcal{C}([0,T];\mathbb{R}) f ∈ C ([ 0 , T ] ; R ) and ∣ ε ∣ ∥ f ∥ ∞ < 1 \lvert\varepsilon\rvert\lVert f\rVert_\infty< 1 ∣ ε ∣ ∥ f ∥ ∞ < 1 . Let V λ ( ε ) V_\lambda(\varepsilon) V λ ( ε ) be the value with a a a replaced by a ( 1 + ε f ) a(1+\varepsilon f) a ( 1 + ε f ) . Testing with ϕ λ \phi^\lambda ϕ λ gives
V λ ( ε ) − V λ ( 0 ) ≤ o ( λ α H ) + ε λ α H ∫ [ 0 , T ] f d μ λ . V_\lambda(\varepsilon)-V_\lambda(0) \le o(\lambda^{\alpha_H})+\varepsilon\lambda^{\alpha_H}\int_{[0,T]} f\,d\mu_\lambda. V λ ( ε ) − V λ ( 0 ) ≤ o ( λ α H ) + ε λ α H ∫ [ 0 , T ] f d μ λ .
The perturbed coefficients satisfy the same assumptions, so
λ − α H V λ ( ε ) → λ → 0 K ( ε ) , \lambda^{-\alpha_H}V_\lambda(\varepsilon)\xrightarrow[\lambda\to0]{}K(\varepsilon), λ − α H V λ ( ε ) λ → 0 K ( ε ) ,
where
K : { { ε ∈ R : ∣ ε ∣ ∥ f ∥ ∞ < 1 } ⟶ R ε ⟼ κ H ∫ 0 T g H ( t ) ( 1 + ε f ( t ) ) p H d t K:
\begin{cases}
\{\varepsilon\in\mathbb{R}:\lvert\varepsilon\rvert\lVert f\rVert_\infty<1\}\longrightarrow\mathbb{R}\\
\varepsilon\longmapsto\kappa_H\int_0^T g_H(t)(1+\varepsilon f(t))^{p_H}\,dt
\end{cases} K : { { ε ∈ R : ∣ ε ∣ ∥ f ∥ ∞ < 1 } ⟶ R ε ⟼ κ H ∫ 0 T g H ( t ) ( 1 + ε f ( t ) ) p H d t
For fixed ε \varepsilon ε , division by ε λ α H \varepsilon\lambda^{\alpha_H} ε λ α H gives the following lower and upper bounds. For ε > 0 \varepsilon>0 ε > 0 ,
K ( ε ) − K ( 0 ) ε ≤ lim inf λ → 0 ∫ [ 0 , T ] f d μ λ , \frac{K(\varepsilon)-K(0)}{\varepsilon}\le\liminf_{\lambda\to0}\int_{[0,T]}f\,d\mu_\lambda, ε K ( ε ) − K ( 0 ) ≤ λ → 0 lim inf ∫ [ 0 , T ] f d μ λ ,
and, for ε < 0 \varepsilon<0 ε < 0 ,
lim sup λ → 0 ∫ [ 0 , T ] f d μ λ ≤ K ( ε ) − K ( 0 ) ε . \limsup_{\lambda\to0}\int_{[0,T]}f\,d\mu_\lambda\le\frac{K(\varepsilon)-K(0)}{\varepsilon}. λ → 0 lim sup ∫ [ 0 , T ] f d μ λ ≤ ε K ( ε ) − K ( 0 ) .
For ε \varepsilon ε near zero, differentiation under the integral gives:
K ′ ( ε ) = p H κ H ∫ 0 T f ( t ) g H ( t ) ( 1 + ε f ( t ) ) p H − 1 d t . K'(\varepsilon)=p_H\kappa_H\int_0^T f(t)g_H(t)(1+\varepsilon f(t))^{p_H-1}\,dt. K ′ ( ε ) = p H κ H ∫ 0 T f ( t ) g H ( t ) ( 1 + ε f ( t ) ) p H − 1 d t .
Then,
K ′ ( 0 ) = p H κ H ∫ 0 T f ( t ) g H ( t ) d t . K'(0)=p_H\kappa_H\int_0^T f(t)g_H(t)\,dt. K ′ ( 0 ) = p H κ H ∫ 0 T f ( t ) g H ( t ) d t .
Letting ε → 0 \varepsilon\to0 ε → 0 from either side proves the tracking limit. For the perturbation b ( 1 + ε f ) b(1+\varepsilon f) b ( 1 + ε f ) , the test term is ε λ α H ∫ [ 0 , T ] f d ν λ \varepsilon\lambda^{\alpha_H}\int_{[0,T]}f\,d\nu_\lambda ε λ α H ∫ [ 0 , T ] f d ν λ and the limiting value is
κ H ∫ 0 T g H ( t ) ( 1 + ε f ( t ) ) α H d t . \kappa_H\int_0^T g_H(t)(1+\varepsilon f(t))^{\alpha_H}\,dt. κ H ∫ 0 T g H ( t ) ( 1 + ε f ( t ) ) α H d t .
The same argument, with exponent α H \alpha_H α H in place of p H p_H p H , proves
∫ [ 0 , T ] f d ν λ → λ → 0 α H κ H ∫ 0 T f ( t ) g H ( t ) d t . \int_{[0,T]}f\,d\nu_\lambda\xrightarrow[\lambda\to0]{}\alpha_H\kappa_H\int_0^T f(t)g_H(t)\,dt. ∫ [ 0 , T ] f d ν λ λ → 0 α H κ H ∫ 0 T f ( t ) g H ( t ) d t .
Taking f ≡ 1 f\equiv1 f ≡ 1 gives the total cost fractions.
Remark 6.7. For Brownian targets the split is 1/3 for tracking and 2/3 for variation. Formula (58) also rules out a leading-order cost atom at either endpoint for near-optimal controls.
A Gaussian Volterra class
Fix H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) and let W W W be a standard Brownian motion on [ 0 , T ] [0,T] [ 0 , T ] , with augmented filtration F t : = F t W \mathcal{F}_t:=\mathcal{F}^W_t F t := F t W . For F , q : [ 0 , T ] → R F,q:[0,T]\to\mathbb{R} F , q : [ 0 , T ] → R and K : ( 0 , T ] → R K:(0,T]\to\mathbb{R} K : ( 0 , T ] → R , define
Θ : { [ 0 , T ] ⟶ L 2 ( Ω ) t ⟼ F ( t ) + ∫ 0 t K ( t − s ) q ( s ) d W s (59) \Theta:
\begin{cases}
[0,T]\longrightarrow L^2(\Omega)\\
t\longmapsto F(t)+\int_0^t K(t-s)q(s)\,dW_s
\end{cases} \tag*{(59)} Θ : { [ 0 , T ] ⟶ L 2 ( Ω ) t ⟼ F ( t ) + ∫ 0 t K ( t − s ) q ( s ) d W s (59)
Suppose that q ∈ C ( [ 0 , T ] ; R ) q \in\mathcal{C}([0,T];\mathbb{R}) q ∈ C ([ 0 , T ] ; R ) and
0 < q − ≤ ∣ q ( t ) ∣ ≤ q + ( 0 ≤ t ≤ T ) . (60) 0<q_{-}\leq|q(t)|\leq q_{+}\qquad(0\leq t\leq T). \tag*{(60)} 0 < q − ≤ ∣ q ( t ) ∣ ≤ q + ( 0 ≤ t ≤ T ) . (60)
The drift is assumed to satisfy F ∈ C α F ( [ 0 , T ] ; R ) F\in\mathcal{C}^{\alpha_F}([0,T];\mathbb{R}) F ∈ C α F ([ 0 , T ] ; R ) for some α F ∈ ( H , 1 ] \alpha_F\in(H,1] α F ∈ ( H , 1 ] . For the kernel, assume K ∈ C 1 ( ( 0 , T ] ; R ) ∩ L 2 ( ( 0 , T ) ) K\in\mathcal{C}^{1}((0,T];\mathbb{R})\cap L^{2}((0,T)) K ∈ C 1 (( 0 , T ] ; R ) ∩ L 2 (( 0 , T )) and
K ( u ) = c K u H − 1 / 2 ( 1 + o ( 1 ) ) as u → 0 , c K ≠ 0. (61) K(u)=c_Ku^{H-1/2}(1+o(1))\quad\text{as }u\to0,\qquad c_K\ne0. \tag*{(61)} K ( u ) = c K u H − 1/2 ( 1 + o ( 1 )) as u → 0 , c K = 0. (61)
Finally, require the bounds
∣ K ( u ) ∣ ≤ C u H − 1 / 2 and ∣ K ′ ( u ) ∣ ≤ C u H − 3 / 2 ( 0 < u ≤ T ) . (62) |K(u)|\leq Cu^{H-1/2}\qquad\text{and}\qquad|K'(u)|\leq Cu^{H-3/2}\qquad(0<u\leq T). \tag*{(62)} ∣ K ( u ) ∣ ≤ C u H − 1/2 and ∣ K ′ ( u ) ∣ ≤ C u H − 3/2 ( 0 < u ≤ T ) . (62)
Set
Γ H : = 1 2 H + ∫ 0 ∞ ( ( 1 + z ) H − 1 / 2 − z H − 1 / 2 ) 2 d z . \Gamma_H:=\frac{1}{2H}+\int_0^\infty\left((1+z)^{H-1/2}-z^{H-1/2}\right)^2\,\mathrm{d}z. Γ H := 2 H 1 + ∫ 0 ∞ ( ( 1 + z ) H − 1/2 − z H − 1/2 ) 2 d z .
With this normalization, define the tangent amplitude by
ρ : { [ 0 , T ] ⟶ ( 0 , ∞ ) t ⟼ ∣ c K q ( t ) ∣ Γ H . (63) \rho:\begin{cases}
[0,T]\longrightarrow(0,\infty)\\
t\longmapsto|c_Kq(t)|\sqrt{\Gamma_H}.
\end{cases} \tag*{(63)} ρ : { [ 0 , T ] ⟶ ( 0 , ∞ ) t ⟼ ∣ c K q ( t ) ∣ Γ H . (63)
The integral defining Γ H \Gamma_H Γ H is finite because its integrand is O ( 1 + z 2 H − 1 ) O(1+z^{2H-1}) O ( 1 + z 2 H − 1 ) at zero and O ( z 2 H − 3 ) O(z^{2H-3}) O ( z 2 H − 3 ) at infinity. When H = 1 / 2 H=1/2 H = 1/2 , the integrand is set equal to zero at every z > 0 z>0 z > 0 . Thus Γ 1 / 2 = 1 \Gamma_{1/2}=1 Γ 1/2 = 1 . Put β : = H − 1 / 2 \beta:=H-1/2 β := H − 1/2 . The Mandelbrot–Van Ness kernel [33 ] is
g u H : { R ⟶ R z ⟼ ( u − z ) + β − ( − z ) + β , u ∈ R . (64) g_u^H:\begin{cases}
\mathbb{R}\longrightarrow\mathbb{R}\\
z\longmapsto(u-z)_+^\beta-(-z)_+^\beta,
\end{cases}
\qquad u\in\mathbb{R}. \tag*{(64)} g u H : { R ⟶ R z ⟼ ( u − z ) + β − ( − z ) + β , u ∈ R . (64)
Here x + β = x β x_+^\beta=x^\beta x + β = x β for x > 0 x>0 x > 0 and is zero for x ≤ 0 x\leq0 x ≤ 0 , including x + 0 = 1 { x > 0 } x_+^0=\mathbf{1}_{\{x>0\}} x + 0 = 1 { x > 0 } . Thus, for H = 1 / 2 H=1/2 H = 1/2 ,
g u 1 / 2 = { 1 ( 0 , u ) , u > 0 , 0 , u = 0 , − 1 ( u , 0 ) , u < 0 , g_u^{1/2}=
\begin{cases}
\mathbf{1}_{(0,u)},&u>0,\\
0,&u=0,\\
-\mathbf{1}_{(u,0)},&u<0,
\end{cases} g u 1/2 = ⎩ ⎨ ⎧ 1 ( 0 , u ) , 0 , − 1 ( u , 0 ) , u > 0 , u = 0 , u < 0 ,
as elements of L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) .
Proposition 6.8 (filtered tangent for Gaussian Volterra targets). For every H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) , the process in (59) –(61) has a continuous adapted modification satisfying assumption 6.1 with ρ \rho ρ in (63) . Therefore,
λ − 2 H / ( 1 + H ) V λ ( Θ ) → λ → 0 κ H ∫ 0 T a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ∣ c K q ( t ) ∣ 2 / ( 1 + H ) Γ H 1 / ( 1 + H ) d t . (65) \lambda^{-2H/(1+H)}V_\lambda(\Theta)\xrightarrow[\lambda\to0]{\kappa_H}\int_0^T a_t^{(1-H)/(1+H)}b_t^{2H/(1+H)}|c_Kq(t)|^{2/(1+H)}\Gamma_H^{1/(1+H)}\,\mathrm{d}t. \tag*{(65)} λ − 2 H / ( 1 + H ) V λ ( Θ ) κ H λ → 0 ∫ 0 T a t ( 1 − H ) / ( 1 + H ) b t 2 H / ( 1 + H ) ∣ c K q ( t ) ∣ 2/ ( 1 + H ) Γ H 1/ ( 1 + H ) d t . (65) The coupling extends the driving Brownian motion by an independent past and applies the Mandelbrot–Van Ness kernel. The kernel estimates then give the uniform L 2 L^2 L 2 approximation and the global increment bound.
Proof. See section H.4 .
Corollary 6.9 (Hölder drifts do not affect the leading term). In the class of proposition 6.8, the leading asymptotic formula (65) does not depend on the drift F F F , if its Hölder exponent is larger than H H H .
Proof. For every fixed R < ∞ R<\infty R < ∞ ,
sup 0 ≤ u ≤ R h − H ∣ F ( t + h u ) − F ( t ) ∣ ≤ [ F ] α F R α F h α F − H → h → 0 0. \sup_{0\leq u\leq R}h^{-H}|F(t+hu)-F(t)|\leq[F]_{\alpha_F}R^{\alpha_F}h^{\alpha_F-H}\xrightarrow[h\to0]{}0. 0 ≤ u ≤ R sup h − H ∣ F ( t + h u ) − F ( t ) ∣ ≤ [ F ] α F R α F h α F − H h → 0 0.
Thus F F F vanishes from the filtered tangent and does not enter ρ \rho ρ or the limiting density.
Remark 6.10. At t = 0 t=0 t = 0 the tangent is generally Riemann–Liouville rather than stationary fBm. The endpoint is handled by the short-interval estimate (120) : an interval of length η \eta η contributes at most C η C\eta C η to the normalized upper bound, uniformly as λ → 0 \lambda\to0 λ → 0 .
Bounds and properties of the fractional cell constant
A causal dual lower bound
Proposition 7.1 (causal dual lower bound). Fix L > 0 L > 0 L > 0 and a complete filtered probability space ( Ω , F , F , P ) (\Omega, \mathcal{F}, \mathbf{F}, \mathbb{P}) ( Ω , F , F , P ) with F = ( F t ) 0 ≤ t ≤ L \mathbf{F} = (\mathcal{F}_t)_{0 \le t \le L} F = ( F t ) 0 ≤ t ≤ L . Let X X X be a continuous, F \mathbf{F} F -adapted real process on [ 0 , L ] [0,L] [ 0 , L ] , satisfying X 0 = 0 X_0 = 0 X 0 = 0 and the integrability assumptions required for U L ( X ; F ) \mathcal{U}_L(X;\mathbf{F}) U L ( X ; F ) . Let W 1 , 2 ( 0 , L ) W^{1,2}(0,L) W 1 , 2 ( 0 , L ) denote the space of functions f ∈ L 2 ( 0 , L ) f \in L^2(0,L) f ∈ L 2 ( 0 , L ) with weak derivative f ′ ∈ L 2 ( 0 , L ) f' \in L^2(0,L) f ′ ∈ L 2 ( 0 , L ) , and let p ∈ L 2 ( Ω ; W 1 , 2 ( 0 , L ) ) p \in L^2(\Omega; W^{1,2}(0,L)) p ∈ L 2 ( Ω ; W 1 , 2 ( 0 , L )) have an almost surely absolutely continuous version with ∣ p t ∣ ≤ 1 |p_t| \le1 ∣ p t ∣ ≤ 1 . The field p p p may be nonadapted, and p ˙ ∈ L 2 ( Ω × [ 0 , L ] , P ⊗ d t ) \dot{p} \in L^2(\Omega\times[0,L], \mathbb{P} \otimes dt) p ˙ ∈ L 2 ( Ω × [ 0 , L ] , P ⊗ d t ) . Write
L p r o g 2 ( Ω × [ 0 , L ] ) : = L 2 ( Ω × [ 0 , L ] , Prog ( F ) , P ⊗ d t ) , L^2_{\mathrm{prog}}(\Omega\times[0,L]) := L^2(\Omega\times[0,L], \operatorname{Prog}(\mathbf{F}), \mathbb{P} \otimes dt), L prog 2 ( Ω × [ 0 , L ]) := L 2 ( Ω × [ 0 , L ] , Prog ( F ) , P ⊗ d t ) , where Prog ( F ) \operatorname{Prog}(\mathbf{F}) Prog ( F ) is the progressive sigma-field of F \mathbf{F} F . Let m m m be the orthogonal projection of p ˙ \dot{p} p ˙ onto this closed subspace. Equivalently, after choosing a jointly measurable version,
m t = E [ p ˙ t ∣ F t ] for P ⊗ d t -a.e. ( ω , t ) . (66) m_t = \mathbb{E}[\dot{p}_t \mid\mathcal{F}_t] \quad\text{for } \mathbb{P} \otimes dt\text{-a.e. }(\omega,t). \tag*{(66)} m t = E [ p ˙ t ∣ F t ] for P ⊗ d t -a.e. ( ω , t ) . (66) Then m ∈ L p r o g 2 ( Ω × [ 0 , L ] ) m \in L^2_{\mathrm{prog}}(\Omega\times[0,L]) m ∈ L prog 2 ( Ω × [ 0 , L ]) and
U L ( X ; F ) ≥ E [ p L X L − ∫ 0 L X t m t d t − 1 2 ∫ 0 L m t 2 d t ] . (67) \mathcal{U}_L(X;\mathbf{F}) \ge\mathbb{E}\left[p_L X_L - \int_0^L X_t m_t\,dt - \frac{1}{2}\int_0^L m_t^2\,dt\right]. \tag*{(67)} U L ( X ; F ) ≥ E [ p L X L − ∫ 0 L X t m t d t − 2 1 ∫ 0 L m t 2 d t ] . (67)
Proof. For ψ ∈ A L ( F ) \psi\in\mathcal{A}_L(\mathbf{F}) ψ ∈ A L ( F ) , the polar inequality and ∣ p ∣ ≤ 1 |p| \le1 ∣ p ∣ ≤ 1 give
TV [ 0 , L ] a u g ( ψ ) ≥ ∫ [ 0 , L ] p t d ψ t , ∣ ψ L − X L ∣ ≥ − p L ( ψ L − X L ) . \operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi) \ge\int_{[0,L]} p_t\,d\psi_t,\qquad|\psi_L-X_L| \ge-p_L(\psi_L-X_L). TV [ 0 , L ] aug ( ψ ) ≥ ∫ [ 0 , L ] p t d ψ t , ∣ ψ L − X L ∣ ≥ − p L ( ψ L − X L ) .
Integration by parts, including the initial atom ψ 0 − ψ 0 − \psi_0-\psi_{0-} ψ 0 − ψ 0 − , uses ψ 0 − = 0 \psi_{0-}=0 ψ 0 − = 0 and continuity of p p p :
∫ [ 0 , L ] p t d ψ t = p 0 ψ 0 + ∫ ( 0 , L ] p t d ψ t = p 0 ψ 0 + p L ψ L − p 0 ψ 0 − ∫ 0 L ψ t p ˙ t d t = p L ψ L − ∫ 0 L ψ t p ˙ t d t . \begin{aligned}
\int_{[0,L]} p_t\,d\psi_t = p_0\psi_0+\int_{(0,L]}p_t\,d\psi_t \\
&= p_0\psi_0+p_L\psi_L-p_0\psi_0-\int_0^L\psi_t\dot{p}_t\,dt \\
&= p_L\psi_L-\int_0^L\psi_t\dot{p}_t\,dt.
\end{aligned} ∫ [ 0 , L ] p t d ψ t = p 0 ψ 0 + ∫ ( 0 , L ] p t d ψ t = p 0 ψ 0 + p L ψ L − p 0 ψ 0 − ∫ 0 L ψ t p ˙ t d t = p L ψ L − ∫ 0 L ψ t p ˙ t d t .
The adapted càdlàg process ψ \psi ψ is progressive and belongs to L 2 ( Ω × [ 0 , L ] ) L^2(\Omega\times[0,L]) L 2 ( Ω × [ 0 , L ]) . Orthogonality of the projection therefore gives
E [ ∫ 0 L ψ t ( p ˙ t − m t ) d t ] = 0. \mathbb{E}\left[\int_0^L \psi_t(\dot{p}_t-m_t)\,dt\right]=0. E [ ∫ 0 L ψ t ( p ˙ t − m t ) d t ] = 0.
The terminal terms cancel except for p L X L p_LX_L p L X L , and completion of the square yields
J L X ( ψ ) ≥ E [ p L X L + ∫ 0 L { 1 2 ( ψ t − X t ) 2 − ψ t m t } d t ] = E [ p L X L − ∫ 0 L X t m t d t − 1 2 ∫ 0 L m t 2 d t ] + 1 2 E [ ∫ 0 L ( ψ t − X t − m t ) 2 d t ] ≥ E [ p L X L − ∫ 0 L X t m t d t − 1 2 ∫ 0 L m t 2 d t ] . \begin{aligned}
\mathcal{J}_L^X(\psi) \ge\mathbb{E}\left[p_LX_L+\int_0^L\left\{\frac{1}{2}(\psi_t-X_t)^2-\psi_t m_t\right\}\,dt\right] \\
&= \mathbb{E}\left[p_LX_L-\int_0^L X_t m_t\,dt-\frac{1}{2}\int_0^L m_t^2\,dt\right]+\frac{1}{2}\mathbb{E}\left[\int_0^L(\psi_t-X_t-m_t)^2\,dt\right] \\
&\ge\mathbb{E}\left[p_LX_L-\int_0^L X_t m_t\,dt-\frac{1}{2}\int_0^L m_t^2\,dt\right].
\end{aligned} J L X ( ψ ) ≥ E [ p L X L + ∫ 0 L { 2 1 ( ψ t − X t ) 2 − ψ t m t } d t ] = E [ p L X L − ∫ 0 L X t m t d t − 2 1 ∫ 0 L m t 2 d t ] + 2 1 E [ ∫ 0 L ( ψ t − X t − m t ) 2 d t ] ≥ E [ p L X L − ∫ 0 L X t m t d t − 2 1 ∫ 0 L m t 2 d t ] .
Taking infima proves (67) .
Remark 7.2. If p p p is adapted and p ˙ \dot{p} p ˙ is progressive, then m = p ˙ m=\dot{p} m = p ˙ . If p p p is not adapted, Jensen gives E [ m t 2 ] ≤ E [ p ˙ t 2 ] \mathbb{E}[m_t^2] \le\mathbb{E}[\dot{p}_t^2] E [ m t 2 ] ≤ E [ p ˙ t 2 ] and E [ X t m t ] = E [ X t p ˙ t ] \mathbb{E}[X_t m_t]=\mathbb{E}[X_t\dot{p}_t] E [ X t m t ] = E [ X t p ˙ t ] .
Corollary 7.3 (ergodic causal dual bound). On the two-sided fBm space of definition 5.1, let P L H \mathcal{P}_L^H P L H be the set of all p ∈ L 2 ( Ω ; W 1 , 2 ( 0 , L ) ) p \in L^2(\Omega; W^{1,2}(0,L)) p ∈ L 2 ( Ω ; W 1 , 2 ( 0 , L )) with an almost surely absolutely continuous version satisfying ∣ p t ∣ ≤ 1 |p_t| \le1 ∣ p t ∣ ≤ 1 for every t ∈ [ 0 , L ] t \in[0,L] t ∈ [ 0 , L ] . For each such p p p , put m t p : = E [ p ˙ t ∣ F t B ] m_t^p := \mathbb{E}[\dot{p}_t \mid\mathcal{F}_t^B] m t p := E [ p ˙ t ∣ F t B ] and define the functional
D H , L : P L H ⟶ R p ⟼ E [ p L B L H − ∫ 0 L ( B t H m t p + 1 2 ( m t p ) 2 ) d t ] . \mathcal{D}_{H,L}:
\begin{array}{rcl}
\mathcal{P}_L^H &\longrightarrow& \mathbb{R}\\
p &\longmapsto& \mathbb{E}\left[p_L B_L^H-\int_0^L\left(B_t^H m_t^p+\frac{1}{2}(m_t^p)^2\right)\,dt\right].
\end{array} D H , L : P L H p ⟶ ⟼ R E [ p L B L H − ∫ 0 L ( B t H m t p + 2 1 ( m t p ) 2 ) d t ] .
Then, for every L > 0 L > 0 L > 0 ,
u H p a s t ( L ) ≥ sup p ∈ P L H D H , L ( p ) , (68) u^{\mathrm{past}}_{H}(L) \ge\sup_{p\in\mathcal{P}^{H}_{L}} \mathcal{D}_{H,L}(p), \tag*{(68)} u H past ( L ) ≥ p ∈ P L H sup D H , L ( p ) , (68)
and
κ H ≥ lim sup L → ∞ 1 L sup p ∈ P L H D H , L ( p ) . (69) \kappa_{H} \ge\limsup_{L\to\infty} \frac{1}{L}\sup_{p\in\mathcal{P}^{H}_{L}} \mathcal{D}_{H,L}(p). \tag*{(69)} κ H ≥ L → ∞ lim sup L 1 p ∈ P L H sup D H , L ( p ) . (69)
Proof. Applying proposition 7.1 with ( X , F ) = ( B H , F B ) (X,\mathbf{F})=(\mathbf{B}^{H},\mathbf{F}^{\mathbf{B}}) ( X , F ) = ( B H , F B ) gives D H , L ( p ) ≤ u H p a s t ( L ) \mathcal{D}_{H,L}(p)\le u^{\mathrm{past}}_{H}(L) D H , L ( p ) ≤ u H past ( L ) for every p ∈ P L H p\in\mathcal{P}^{H}_{L} p ∈ P L H . Taking the supremum proves the finite-horizon claim and, after division by L L L , yields
1 L sup p ∈ P L H D H , L ( p ) ≤ u H p a s t ( L ) L . \frac{1}{L}\sup_{p\in\mathcal{P}^{H}_{L}}\mathcal{D}_{H,L}(p)\le\frac{u^{\mathrm{past}}_{H}(L)}{L}. L 1 p ∈ P L H sup D H , L ( p ) ≤ L u H past ( L ) .
The upper limit and theorem 5.3 then give the ergodic bound.
Causal linear-filter bounds
Proposition 7.4 (single-window bounds). For every H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) ,
C l o w ( H ) ≤ κ H ≤ C m a ( H ) , (70) C_{\mathrm{low}}(H)\le\kappa_{H}\le C_{\mathrm{ma}}(H), \tag*{(70)} C low ( H ) ≤ κ H ≤ C ma ( H ) , (70)
where
C l o w : { ( 0 , 1 ) ⟶ ( 0 , ∞ ) H ⟼ m 1 ( 1 + H ) 2 ( m 1 ( 1 − H ) 4 ) ( 1 − H ) / ( 1 + H ) (71) C_{\mathrm{low}}:\begin{cases}
(0,1)\longrightarrow(0,\infty)\\
H\longmapsto\dfrac{m_{1}(1+H)}{2}\left(\dfrac{m_{1}(1-H)}{4}\right)^{(1-H)/(1+H)}
\end{cases} \tag*{(71)} C low : ⎩ ⎨ ⎧ ( 0 , 1 ) ⟶ ( 0 , ∞ ) H ⟼ 2 m 1 ( 1 + H ) ( 4 m 1 ( 1 − H ) ) ( 1 − H ) / ( 1 + H ) (71)
and
C m a : { ( 0 , 1 ) ⟶ ( 0 , ∞ ) H ⟼ 1 4 ( 1 − H ) ( 2 m 1 ( 1 − H ) ( 1 + H ) H ) 2 H / ( 1 + H ) (72) C_{\mathrm{ma}}:\begin{cases}
(0,1)\longrightarrow(0,\infty)\\
H\longmapsto\dfrac{1}{4(1-H)}\left(\dfrac{2m_{1}(1-H)(1+H)}{H}\right)^{2H/(1+H)}
\end{cases} \tag*{(72)} C ma : ⎩ ⎨ ⎧ ( 0 , 1 ) ⟶ ( 0 , ∞ ) H ⟼ 4 ( 1 − H ) 1 ( H 2 m 1 ( 1 − H ) ( 1 + H ) ) 2 H / ( 1 + H ) (72)
Proof. See § 18 .§ 1 .
Theorem 7.5 (continuity of the fractional cell constant). The map H ↦ κ H H\mapsto\kappa_{H} H ↦ κ H is continuous on ( 0 , 1 ) (0,1) ( 0 , 1 ) . For each fixed L > 0 L>0 L > 0 , H ↦ u H p a s t ( L ) H\mapsto u^{\mathrm{past}}_{H}(L) H ↦ u H past ( L ) is also continuous .
Proof. Fix H 0 ∈ ( 0 , 1 ) H_{0}\in(0,1) H 0 ∈ ( 0 , 1 ) and a compact interval I = [ H − , H + ] ⊂ ( 0 , 1 ) I=[H_{-},H_{+}]\subset(0,1) I = [ H − , H + ] ⊂ ( 0 , 1 ) containing H 0 H_{0} H 0 in its interior. Couple all H ∈ I H\in I H ∈ I on one two-sided Brownian space by
B t H : = Γ H − 1 / 2 ∫ R g t H ( s ) d W s . B^{H}_{t}:=\Gamma_{H}^{-1/2}\int_{\mathbb{R}}g^{H}_{t}(s)\,\mathrm{d}W_{s}. B t H := Γ H − 1/2 ∫ R g t H ( s ) d W s .
For fixed t t t , the kernels converge pointwise almost everywhere as H → H 0 H\to H_{0} H → H 0 . Near s = 0 , t s=0,t s = 0 , t , their squares have an integrable envelope of the form
C t ( 1 + ∣ s ∣ 2 H − − 1 + ∣ s − t ∣ 2 H − − 1 ) , 2 H − − 1 > − 1. C_{t}\left(1+|s|^{2H_{-}-1}+|s-t|^{2H_{-}-1}\right),\qquad2H_{-}-1>-1. C t ( 1 + ∣ s ∣ 2 H − − 1 + ∣ s − t ∣ 2 H − − 1 ) , 2 H − − 1 > − 1.
For sufficiently negative s s s , they are bounded by C t ∣ s ∣ 2 H + − 3 C_{t}|s|^{2H_{+}-3} C t ∣ s ∣ 2 H + − 3 , with 2 H + − 3 < − 1 2H_{+}-3<-1 2 H + − 3 < − 1 . Dominated convergence gives
∥ g t H − g t H 0 ∥ 2 → H → H 0 0 , Γ H = ∥ g t H ∥ 2 2 → H → H 0 Γ H 0 > 0. \lVert g^{H}_{t}-g^{H_{0}}_{t}\rVert_{2}\xrightarrow[H\to H_{0}]{}0,\qquad\Gamma_{H}=\lVert g^{H}_{t}\rVert_{2}^{2}\xrightarrow[H\to H_{0}]{}\Gamma_{H_{0}}>0. ∥ g t H − g t H 0 ∥ 2 H → H 0 0 , Γ H = ∥ g t H ∥ 2 2 H → H 0 Γ H 0 > 0.
The Wiener isometry and the triangle inequality therefore give
∥ B t H − B t H 0 ∥ L 2 = ∥ Γ H − 1 / 2 g t H − Γ H 0 − 1 / 2 g t H 0 ∥ 2 ≤ Γ H − 1 / 2 ∥ g t H − g t H 0 ∥ 2 + ∣ Γ H − 1 / 2 − Γ H 0 − 1 / 2 ∣ ∥ g t H 0 ∥ 2 → H → H 0 0. \begin{aligned}
\lVert B^{H}_{t}-B^{H_{0}}_{t}\rVert_{L^{2}}
&=\lVert\Gamma_{H}^{-1/2}g^{H}_{t}-\Gamma_{H_{0}}^{-1/2}g^{H_{0}}_{t}\rVert_{2}\\
&\le\Gamma_{H}^{-1/2}\lVert g^{H}_{t}-g^{H_{0}}_{t}\rVert_{2}
+\left|\Gamma_{H}^{-1/2}-\Gamma_{H_{0}}^{-1/2}\right|\lVert g^{H_{0}}_{t}\rVert_{2}
\xrightarrow[H\to H_{0}]{}0.
\end{aligned} ∥ B t H − B t H 0 ∥ L 2 = ∥ Γ H − 1/2 g t H − Γ H 0 − 1/2 g t H 0 ∥ 2 ≤ Γ H − 1/2 ∥ g t H − g t H 0 ∥ 2 + Γ H − 1/2 − Γ H 0 − 1/2 ∥ g t H 0 ∥ 2 H → H 0 0.
On [ 0 , L ] [0,L] [ 0 , L ] ,
E [ ∣ B t H − B t H 0 ∣ 2 ] ≤ 2 t 2 H + 2 t 2 H 0 ≤ 4 ( t 2 H − 1 { t ≤ 1 } + t 2 H + 1 { t > 1 } ) . \mathbb{E}\left[\left|B^{H}_{t}-B^{H_{0}}_{t}\right|^{2}\right]\le2t^{2H}+2t^{2H_{0}}\le4\left(t^{2H_{-}}\mathbf{1}_{\{t\le1\}}+t^{2H_{+}}\mathbf{1}_{\{t>1\}}\right). E [ B t H − B t H 0 2 ] ≤ 2 t 2 H + 2 t 2 H 0 ≤ 4 ( t 2 H − 1 { t ≤ 1 } + t 2 H + 1 { t > 1 } ) .
Another application of dominated convergence yields
∥ B H − B H 0 ∥ L 2 ( Ω × [ 0 , L ] ) → H → H 0 0 , \lVert B^{H}-B^{H_{0}}\rVert_{L^{2}(\Omega\times[0,L])}\xrightarrow[H\to H_{0}]{}0, ∥ B H − B H 0 ∥ L 2 ( Ω × [ 0 , L ]) H → H 0 0 ,
and
E [ ∣ B L H − B L H 0 ∣ ] → H → H 0 0. \mathbb{E}\left[\left|B^{H}_{L}-B^{H_{0}}_{L}\right|\right]\xrightarrow[H\to H_{0}]{}0. E [ B L H − B L H 0 ] H → H 0 0.
By lemma A.1, all coupled targets have the same completed full-past filtration. The common-filtration stability estimate in lemma 6.3 therefore gives
∣ u H p a s t ( L ) − u H 0 p a s t ( L ) ∣ ≤ E [ ∣ B L H − B L H 0 ∣ + 1 2 ∥ B L H − B L H 0 ∥ L 2 ( Ω × [ 0 , L ] ) 2 ] + 2 u H 0 p a s t ( L ) ∥ B L H − B L H 0 ∥ L 2 ( Ω × [ 0 , L ] ) → H → H 0 0. \left|u_H^{\mathrm{past}}(L)-u_{H_0}^{\mathrm{past}}(L)\right|
\leq\mathbb{E}\left[\left|B_L^H-B_L^{H_0}\right|+\frac{1}{2}\left\|B_L^H-B_L^{H_0}\right\|_{L^2(\Omega\times[0,L])}^2\right]
+\sqrt{2u_{H_0}^{\mathrm{past}}(L)}\left\|B_L^H-B_L^{H_0}\right\|_{L^2(\Omega\times[0,L])}
\xrightarrow[H\to H_0]{}0. u H past ( L ) − u H 0 past ( L ) ≤ E [ B L H − B L H 0 + 2 1 B L H − B L H 0 L 2 ( Ω × [ 0 , L ]) 2 ] + 2 u H 0 past ( L ) B L H − B L H 0 L 2 ( Ω × [ 0 , L ]) H → H 0 0.
Finally, theorem 5.4 gives
∣ κ H − κ H 0 ∣ ≤ 2 C 0 L + ∣ u H p a s t ( L ) − u H 0 p a s t ( L ) ∣ L . \left|\kappa_H-\kappa_{H_0}\right|\leq\frac{2C_0}{L}+\frac{\left|u_H^{\mathrm{past}}(L)-u_{H_0}^{\mathrm{past}}(L)\right|}{L}. ∣ κ H − κ H 0 ∣ ≤ L 2 C 0 + L u H past ( L ) − u H 0 past ( L ) .
First let H → H 0 H\to H_0 H → H 0 , then L → ∞ L\to\infty L → ∞ .
Corollary 7.6 (endpoint behavior of κ H \kappa_H κ H ). The fractional cell constant satisfies
lim H ↑ 1 κ H = 2 π . (73) \lim_{H\uparrow1}\kappa_H=\sqrt{\frac{2}{\pi}}. \tag*{(73)} H ↑ 1 lim κ H = π 2 . (73) At the other endpoint,
1 4 π ≤ lim inf H ↓ 0 κ H ≤ lim sup H ↓ 0 κ H ≤ 1 4 . (74) \frac{1}{4\pi}\leq\liminf_{H\downarrow0}\kappa_H\leq\limsup_{H\downarrow0}\kappa_H\leq\frac{1}{4}. \tag*{(74)} 4 π 1 ≤ H ↓ 0 lim inf κ H ≤ H ↓ 0 lim sup κ H ≤ 4 1 . (74)
Proof. Put ε : = 1 − H \varepsilon:=1-H ε := 1 − H . Since ε log ε → 0 \varepsilon\log\varepsilon\to0 ε log ε → 0 , we have
C l o w ( 1 − ε ) = m 1 ( 2 − ε ) 2 exp [ ε 2 − ε log ( m 1 ε 4 ) ] → ε → 0 m 1 , C_{\mathrm{low}}(1-\varepsilon)=\frac{m_1(2-\varepsilon)}{2}\exp\left[\frac{\varepsilon}{2-\varepsilon}\log\left(\frac{m_1\varepsilon}{4}\right)\right]\xrightarrow[\varepsilon\to0]{}m_1, C low ( 1 − ε ) = 2 m 1 ( 2 − ε ) exp [ 2 − ε ε log ( 4 m 1 ε ) ] ε → 0 m 1 ,
and
C m a ( 1 − ε ) = m 1 ( 2 − ε ) 2 ( 1 − ε ) exp [ − ε 2 − ε log ( 2 m 1 ε ( 2 − ε ) 1 − ε ) ] → ε → 0 m 1 . C_{\mathrm{ma}}(1-\varepsilon)=\frac{m_1(2-\varepsilon)}{2(1-\varepsilon)}\exp\left[-\frac{\varepsilon}{2-\varepsilon}\log\left(\frac{2m_1\varepsilon(2-\varepsilon)}{1-\varepsilon}\right)\right]\xrightarrow[\varepsilon\to0]{}m_1. C ma ( 1 − ε ) = 2 ( 1 − ε ) m 1 ( 2 − ε ) exp [ − 2 − ε ε log ( 1 − ε 2 m 1 ε ( 2 − ε ) ) ] ε → 0 m 1 .
At the other endpoint,
C l o w ( H ) → H → 0 m 1 2 m 1 4 = 1 4 π , C_{\mathrm{low}}(H)\xrightarrow[H\to0]{}\frac{m_1}{2}\frac{m_1}{4}=\frac{1}{4\pi}, C low ( H ) H → 0 2 m 1 4 m 1 = 4 π 1 ,
and
log C m a ( H ) = − log ( 4 ( 1 − H ) ) + 2 H 1 + H log ( 2 m 1 ( 1 − H ) ( 1 + H ) H ) → H → 0 − log 4 , \log C_{\mathrm{ma}}(H)=-\log(4(1-H))+\frac{2H}{1+H}\log\left(\frac{2m_1(1-H)(1+H)}{H}\right)\xrightarrow[H\to0]{}-\log4, log C ma ( H ) = − log ( 4 ( 1 − H )) + 1 + H 2 H log ( H 2 m 1 ( 1 − H ) ( 1 + H ) ) H → 0 − log 4 ,
since H log H → H → 0 0 H\log H\xrightarrow[H\to0]{}0 H log H H → 0 0 . The bounds in (70) give both conclusions. □
The moving-average upper bound extends to linear combinations of windows. The resulting bound, its scale-free form, and the optimal weights for fixed windows are given in § 18 .§ 2 .
Finite-dimensional Markov approximations
Assume H ∈ ( 0 , 1 / 2 ) H\in(0,1/2) H ∈ ( 0 , 1/2 ) and put β : = 1 2 − H ∈ ( 0 , 1 / 2 ) \beta:=\frac{1}{2}-H\in(0,1/2) β := 2 1 − H ∈ ( 0 , 1/2 ) . Euler’s gamma function is
Γ : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) z ⟼ ∫ 0 ∞ r z − 1 e − r d r . \begin{aligned}
\Gamma:\quad(0,\infty)&\longrightarrow(0,\infty)\\
z&\longmapsto\int_0^\infty r^{z-1}e^{-r}\,\mathrm{d}r.
\end{aligned} Γ : ( 0 , ∞ ) z ⟶ ( 0 , ∞ ) ⟼ ∫ 0 ∞ r z − 1 e − r d r .
Define the fractional kernel by
k H : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) t ⟼ 2 H t H − 1 / 2 = 2 H Γ ( β ) ∫ 0 ∞ e − x t x β − 1 d x . (75) \begin{aligned}
k_H:\quad(0,\infty)&\longrightarrow(0,\infty)\\
t&\longmapsto\sqrt{2H}\,t^{H-1/2}
=\frac{\sqrt{2H}}{\Gamma(\beta)}\int_0^\infty e^{-xt}x^{\beta-1}\,\mathrm{d}x. \tag*{(75)}
\end{aligned} k H : ( 0 , ∞ ) t ⟶ ( 0 , ∞ ) ⟼ 2 H t H − 1/2 = Γ ( β ) 2 H ∫ 0 ∞ e − x t x β − 1 d x . (75)
The substitution r = x t r=xt r = x t gives
∫ 0 ∞ e − x t x β − 1 d x = t − β Γ ( β ) , Γ ( β + 1 ) = β Γ ( β ) . \int_0^\infty e^{-xt}x^{\beta-1}\,\mathrm{d}x=t^{-\beta}\Gamma(\beta),\qquad\Gamma(\beta+1)=\beta\Gamma(\beta). ∫ 0 ∞ e − x t x β − 1 d x = t − β Γ ( β ) , Γ ( β + 1 ) = β Γ ( β ) .
For a function k k k on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) , define its causal extension by
k ∘ : R ⟶ R r ⟼ { k ( r ) , r > 0 , 0 , r ≤ 0. k^\circ:
\begin{array}{rcl}
\mathbb{R} &\longrightarrow& \mathbb{R}\\
r &\longmapsto&
\begin{cases}
k(r), & r>0,\\
0, & r\le0.
\end{cases}
\end{array} k ∘ : R r ⟶ ⟼ R { k ( r ) , 0 , r > 0 , r ≤ 0.
Following the Ornstein–Uhlenbeck mixture approach [20 , 26 ] , we approximate this kernel by positive exponential sums. For n ∈ N n\in\mathbb{N} n ∈ N , let N n ∈ N N_n\in\mathbb{N} N n ∈ N and define
k n : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) t ⟼ ∑ i = 1 N n w i , n e − x i , n t , w i , n , x i , n > 0. (76) k_n:
\begin{array}{rcl}
(0,\infty) &\longrightarrow& (0,\infty)\\
t &\longmapsto& \sum_{i=1}^{N_n}w_{i,n}e^{-x_{i,n}t},
\end{array}
\qquad w_{i,n},x_{i,n}>0. \tag*{(76)} k n : ( 0 , ∞ ) t ⟶ ⟼ ( 0 , ∞ ) ∑ i = 1 N n w i , n e − x i , n t , w i , n , x i , n > 0. (76)
On a two-sided Brownian space, define
A H : = 2 H Γ H , Y t i , n : = ∫ − ∞ t e − x i , n ( t − s ) d W s , A_H:=\sqrt{2H\Gamma_H},\qquad
Y_t^{i,n}:=\int_{-\infty}^{t}e^{-x_{i,n}(t-s)}\,\mathrm{d}W_s, A H := 2 H Γ H , Y t i , n := ∫ − ∞ t e − x i , n ( t − s ) d W s ,
and
Z t n : = 1 A H ∑ i = 1 N n w i , n ( Y t i , n − Y 0 i , n ) = 1 A H ∫ R { k n ∘ ( t − s ) − k n ∘ ( − s ) } d W s . (77) \begin{aligned}
Z_t^n:=\frac{1}{A_H}\sum_{i=1}^{N_n}w_{i,n}\left(Y_t^{i,n}-Y_0^{i,n}\right) \\
&=\frac{1}{A_H}\int_{\mathbb{R}}\left\{k_n^\circ(t-s)-k_n^\circ(-s)\right\}\,\mathrm{d}W_s. \tag*{(77)}
\end{aligned} Z t n := A H 1 i = 1 ∑ N n w i , n ( Y t i , n − Y 0 i , n ) = A H 1 ∫ R { k n ∘ ( t − s ) − k n ∘ ( − s ) } d W s . (77)
The coupled two-sided fBm is
B t H : = 1 A H ∫ R { k H ∘ ( t − s ) − k H ∘ ( − s ) } d W s , t ∈ R . (78) B_t^H:=\frac{1}{A_H}\int_{\mathbb{R}}\left\{k_H^\circ(t-s)-k_H^\circ(-s)\right\}\,\mathrm{d}W_s,\qquad t\in\mathbb{R}. \tag*{(78)} B t H := A H 1 ∫ R { k H ∘ ( t − s ) − k H ∘ ( − s ) } d W s , t ∈ R . (78)
Let W 0 = 0 W_0=0 W 0 = 0 and let Y t n : = ( Y t 1 , n , … , Y t N n , n ) ∈ R N n Y_t^n:=(Y_t^{1,n},\ldots,Y_t^{N_n,n})\in\mathbb{R}^{N_n} Y t n := ( Y t 1 , n , … , Y t N n , n ) ∈ R N n . Define the completed full-past filtrations
F t W : = σ ( W s : s ≤ t ) ‾ P , F t Y n : = σ ( Y s n : s ≤ t ) ‾ P , t ∈ R , \mathcal{F}_t^W:=\overline{\sigma(W_s:s\le t)}^{\mathbb{P}},\qquad
\mathcal{F}_t^{Y^n}:=\overline{\sigma(Y_s^n:s\le t)}^{\mathbb{P}},\qquad t\in\mathbb{R}, F t W := σ ( W s : s ≤ t ) P , F t Y n := σ ( Y s n : s ≤ t ) P , t ∈ R ,
and write F W = ( F t W ) t ≥ 0 \mathbb{F}^W=(\mathcal{F}_t^W)_{t\ge0} F W = ( F t W ) t ≥ 0 and F Y n = ( F t Y n ) t ≥ 0 \mathbb{F}^{Y^n}=(\mathcal{F}_t^{Y^n})_{t\ge0} F Y n = ( F t Y n ) t ≥ 0 for control times. The process at negative times represents the past observed at time zero.
Lemma 8.1 (factor and Brownian filtrations are equal). For every t ≥ 0 t\ge0 t ≥ 0 ,
F t Y n = F t W . (79) \mathcal{F}_t^{Y^n}=\mathcal{F}_t^W. \tag*{(79)} F t Y n = F t W . (79)
Therefore, for every $L>0,
U L ( Z n ; F W ) = U L ( Z n ; F Y n ) . (80) \mathcal{U}_L(Z^n;\mathbb{F}^W)=\mathcal{U}_L(Z^n;\mathbb{F}^{Y^n}). \tag*{(80)} U L ( Z n ; F W ) = U L ( Z n ; F Y n ) . (80)
Proof. Each factor is a Wiener integral of the Brownian past, so F t Y n ⊂ F t W \mathcal{F}_t^{Y^n}\subset\mathcal{F}_t^W F t Y n ⊂ F t W for t ≥ 0 t\ge0 t ≥ 0 . Conversely, fix any component i i i . Its OU equation is
d Y s i , n = − x i , n Y s i , n d s + d W s . \mathrm{d}Y_s^{i,n}=-x_{i,n}Y_s^{i,n}\,\mathrm{d}s+\mathrm{d}W_s. d Y s i , n = − x i , n Y s i , n d s + d W s .
Integrating from zero to s ≤ t s\le t s ≤ t gives
W s = Y s i , n − Y 0 i , n + x i , n ∫ 0 s Y r i , n d r . W_s=Y_s^{i,n}-Y_0^{i,n}+x_{i,n}\int_0^sY_r^{i,n}\,\mathrm{d}r. W s = Y s i , n − Y 0 i , n + x i , n ∫ 0 s Y r i , n d r .
Because t ≥ 0 t\ge0 t ≥ 0 , the observed factor past contains the whole interval between s s s and zero, including Y 0 i , n Y_0^{i,n} Y 0 i , n . Thus it recovers every W s W_s W s , s ≤ t s\le t s ≤ t , and the completed filtrations coincide. The control classes, and hence their reset values, are identical. □ \square □
For L > 0 L>0 L > 0 , define
u n : ( 0 , ∞ ) ⟶ [ 0 , ∞ ) L ⟼ U L ( Z n ; F W ) . (81) u_n:
\begin{array}{rcl}
(0,\infty)&\longrightarrow&[0,\infty)\\
L&\longmapsto&\mathcal{U}_L(Z^n;\mathbb{F}^W).
\end{array} \tag*{(81)} u n : ( 0 , ∞ ) L ⟶ ⟼ [ 0 , ∞ ) U L ( Z n ; F W ) . (81)
Theorem 8.2 (quantitative convergence of factor cell values). For n ∈ N n\in\mathbb{N} n ∈ N and L ≥ 1 L\ge1 L ≥ 1 , measure the approximation error by
A n , L : = E [ ∫ 0 L ∣ Z t n − B t H ∣ 2 d t ] , B n , L : = E [ ∣ Z L n − B L H ∣ ] . A_{n,L}:=\mathbb{E}\left[\int_0^L\left|Z_t^n-B_t^H\right|^2\,\mathrm{d}t\right],\qquad
B_{n,L}:=\mathbb{E}\left[\left|Z_L^n-B_L^H\right|\right]. A n , L := E [ ∫ 0 L Z t n − B t H 2 d t ] , B n , L := E [ Z L n − B L H ] .
There is a finite constant C C C , independent of n n n , L L L , such that
∣ u n ( L ) L − κ H ∣ ≤ C L + B n , L L + A n , L 2 L + C A n , L L . (82) \left|\frac{u_n(L)}{L}-\kappa_H\right|\leq\frac{C}{L}+\frac{B_{n,L}}{L}+\frac{A_{n,L}}{2L}+C\sqrt{\frac{A_{n,L}}{L}}. \tag*{(82)} L u n ( L ) − κ H ≤ L C + L B n , L + 2 L A n , L + C L A n , L . (82)
In particular, if n k , L k → k → ∞ ∞ n_k,L_k \xrightarrow[k\to\infty]{} \infty n k , L k k → ∞ ∞ and
1 L k E [ ∫ 0 L k ∣ Z t n k − B t H ∣ 2 d t ] + 1 L k 2 E [ ∣ Z L k n k − B L k H ∣ 2 ] → k → ∞ 0 , (83) \frac{1}{L_k}\mathbb{E}\left[\int_0^{L_k}\left|Z_t^{n_k}-B_t^H\right|^2\,\mathrm{d}t\right]+\frac{1}{L_k^2}\mathbb{E}\left[\left|Z_{L_k}^{n_k}-B_{L_k}^H\right|^2\right]\xrightarrow[k\to\infty]{}0, \tag*{(83)} L k 1 E [ ∫ 0 L k Z t n k − B t H 2 d t ] + L k 2 1 E [ Z L k n k − B L k H 2 ] k → ∞ 0 , (83)
then
u n k ( L k ) L k → k → ∞ κ H . (84) \frac{u_{n_k}(L_k)}{L_k}\xrightarrow[k\to\infty]{}\kappa_H. \tag*{(84)} L k u n k ( L k ) k → ∞ κ H . (84)
Proof. By lemma A.1, the fBm reset value in F W \mathrm{FW} FW is u H p a s t ( L ) u_H^{\mathrm{past}}(L) u H past ( L ) . Stability in this common filtration gives
∣ u n ( L ) − u H p a s t ( L ) ∣ ≤ B n , L + A n , L 2 + 2 u H p a s t ( L ) A n , L . \left|u_n(L)-u_H^{\mathrm{past}}(L)\right|\leq B_{n,L}+\frac{A_{n,L}}{2}+\sqrt{2u_H^{\mathrm{past}}(L)A_{n,L}}. u n ( L ) − u H past ( L ) ≤ B n , L + 2 A n , L + 2 u H past ( L ) A n , L .
By theorem 5.4, for L ≥ 1 L\geq1 L ≥ 1 ,
0 ≤ u H p a s t ( L ) − κ H L ≤ 2 2 κ H , 0\leq u_H^{\mathrm{past}}(L)-\kappa_H L\leq2\sqrt{2\kappa_H}, 0 ≤ u H past ( L ) − κ H L ≤ 2 2 κ H ,
so
u H p a s t ( L ) ≤ ( κ H + 2 2 κ H ) L . u_H^{\mathrm{past}}(L)\leq\left(\kappa_H+2\sqrt{2\kappa_H}\right)L. u H past ( L ) ≤ ( κ H + 2 2 κ H ) L .
The triangle inequality combines these bounds into
∣ u n ( L ) L − κ H ∣ ≤ 2 2 κ H L + B n , L L + A n , L 2 L + 2 ( κ H + 2 2 κ H ) A n , L L . \left|\frac{u_n(L)}{L}-\kappa_H\right|\leq\frac{2\sqrt{2\kappa_H}}{L}+\frac{B_{n,L}}{L}+\frac{A_{n,L}}{2L}+\sqrt{2\left(\kappa_H+2\sqrt{2\kappa_H}\right)}\sqrt{\frac{A_{n,L}}{L}}. L u n ( L ) − κ H ≤ L 2 2 κ H + L B n , L + 2 L A n , L + 2 ( κ H + 2 2 κ H ) L A n , L .
This is (81) with a constant independent of n n n , L L L . Under (82), Cauchy–Schwarz also gives
B n k , L k L k ≤ ( E [ ∣ Z L k n k − B L k H ∣ 2 ] L k 2 ) 1 / 2 → k → ∞ 0 , \frac{B_{n_k,L_k}}{L_k}\leq\left(\frac{\mathbb{E}\left[\left|Z_{L_k}^{n_k}-B_{L_k}^H\right|^2\right]}{L_k^2}\right)^{1/2}\xrightarrow[k\to\infty]{}0, L k B n k , L k ≤ L k 2 E [ Z L k n k − B L k H 2 ] 1/2 k → ∞ 0 ,
and
A n k , L k L k → k → ∞ 0. \frac{A_{n_k,L_k}}{L_k}\xrightarrow[k\to\infty]{}0. L k A n k , L k k → ∞ 0.
Every term in the bound vanishes. □
Corollary 8.3 (kernel criterion). For n ∈ N n\in\mathbb{N} n ∈ N , put δ k n : = k n − k H \delta k_n:=k_n-k_H δ k n := k n − k H on ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and, for t ≥ 0 t\geq0 t ≥ 0 , define
ε n ( t ) 2 : = 1 A H 2 ∫ 0 t ∣ δ k n ( r ) ∣ 2 d r + 1 A H 2 ∫ 0 ∞ ∣ δ k n ( t + r ) − δ k n ( r ) ∣ 2 d r . (85) \varepsilon_n(t)^2:=\frac{1}{A_H^2}\int_0^t|\delta k_n(r)|^2\,\mathrm{d}r+\frac{1}{A_H^2}\int_0^\infty|\delta k_n(t+r)-\delta k_n(r)|^2\,\mathrm{d}r. \tag*{(85)} ε n ( t ) 2 := A H 2 1 ∫ 0 t ∣ δ k n ( r ) ∣ 2 d r + A H 2 1 ∫ 0 ∞ ∣ δ k n ( t + r ) − δ k n ( r ) ∣ 2 d r . (85) Then
E [ ∣ Z t n − B t H ∣ 2 ] = ε n ( t ) 2 . (86) \mathbb{E}\left[\left|Z_t^n-B_t^H\right|^2\right]=\varepsilon_n(t)^2. \tag*{(86)} E [ Z t n − B t H 2 ] = ε n ( t ) 2 . (86) Hence (83) holds for any n k → k → ∞ ∞ n_k\xrightarrow[k\to\infty]{}\infty n k k → ∞ ∞ and L k → k → ∞ ∞ L_k\xrightarrow[k\to\infty]{}\infty L k k → ∞ ∞ such that
1 L k ∫ 0 L k ε n k ( t ) 2 d t + ε n k ( L k ) 2 L k 2 → k → ∞ 0. (87) \frac{1}{L_k}\int_0^{L_k}\varepsilon_{n_k}(t)^2\,\mathrm{d}t+\frac{\varepsilon_{n_k}(L_k)^2}{L_k^2}\xrightarrow[k\to\infty]{}0. \tag*{(87)} L k 1 ∫ 0 L k ε n k ( t ) 2 d t + L k 2 ε n k ( L k ) 2 k → ∞ 0. (87) Proof. The difference is the Wiener integral
Z t n − B t H = 1 A H ∫ R { δ k n ∘ ( t − s ) − δ k n ∘ ( − s ) } d W s . Z_t^n-B_t^H=\frac{1}{A_H}\int_{\mathbb{R}}\{\delta k_n^\circ(t-s)-\delta k_n^\circ(-s)\}\,\mathrm{d}W_s. Z t n − B t H = A H 1 ∫ R { δ k n ∘ ( t − s ) − δ k n ∘ ( − s )} d W s . The integrand vanishes for s ≥ t s\geq t s ≥ t . On 0 < s < t 0<s<t 0 < s < t it equals δ k n ( t − s ) \delta k_n(t-s) δ k n ( t − s ) , whereas on s ≤ 0 s\leq0 s ≤ 0 it equals δ k n ( t − s ) − δ k n ( − s ) \delta k_n(t-s)-\delta k_n(-s) δ k n ( t − s ) − δ k n ( − s ) . The Wiener isometry, followed by the substitutions r = t − s r=t-s r = t − s and r = − s r=-s r = − s in the two integrals, therefore gives
A H 2 E [ ∣ Z t n − B t H ∣ 2 ] = ∫ 0 t ∣ δ k n ( t − s ) ∣ 2 d s + ∫ − ∞ 0 ∣ δ k n ( t − s ) − δ k n ( − s ) ∣ 2 d s A_H^2\mathbb{E}\left[\left|Z_t^n-B_t^H\right|^2\right]=\int_0^t|\delta k_n(t-s)|^2\,\mathrm{d}s+\int_{-\infty}^0|\delta k_n(t-s)-\delta k_n(-s)|^2\,\mathrm{d}s A H 2 E [ Z t n − B t H 2 ] = ∫ 0 t ∣ δ k n ( t − s ) ∣ 2 d s + ∫ − ∞ 0 ∣ δ k n ( t − s ) − δ k n ( − s ) ∣ 2 d s = ∫ 0 t ∣ δ k n ( r ) ∣ 2 d r + ∫ 0 ∞ ∣ δ k n ( t + r ) − δ k n ( r ) ∣ 2 d r =\int_0^t|\delta k_n(r)|^2\,\mathrm{d}r+\int_0^\infty|\delta k_n(t+r)-\delta k_n(r)|^2\,\mathrm{d}r = ∫ 0 t ∣ δ k n ( r ) ∣ 2 d r + ∫ 0 ∞ ∣ δ k n ( t + r ) − δ k n ( r ) ∣ 2 d r = A H 2 ε n ( t ) 2 . =A_H^2\varepsilon_n(t)^2. = A H 2 ε n ( t ) 2 . The second integral controls the stationary past. No global L 2 ( 0 , ∞ ) L^2(0,\infty) L 2 ( 0 , ∞ ) assumption on δ k n \delta k_n δ k n is needed. Substitution in (82) proves the claim. □
Theorem 8.4 (a positive quadrature). Fix H ∈ ( 0 , 1 / 2 ) H \in(0,1/2) H ∈ ( 0 , 1/2 ) and write β = 1 / 2 − H \beta= 1/2-H β = 1/2 − H . Given 0 < ε < R 0<\varepsilon<R 0 < ε < R and 0 < η ≤ 1 0<\eta\le1 0 < η ≤ 1 , choose a partition ε = x 0 < ⋯ < x N = R \varepsilon=x_0<\cdots<x_N=R ε = x 0 < ⋯ < x N = R into intervals [ x j , x j + 1 ] [x_j,x_{j+1}] [ x j , x j + 1 ] , j = 0 , … , N − 1 j=0,\ldots,N-1 j = 0 , … , N − 1 , with x j + 1 / x j ≤ 1 + η x_{j+1}/x_j\le1+\eta x j + 1 / x j ≤ 1 + η . Define
k ε , R , η : ( 0 , ∞ ) ⟶ ( 0 , ∞ ) t ⟼ ∑ j = 0 N − 1 w j e − x j t , (88) \begin{aligned}
k_{\varepsilon,R,\eta}:(0,\infty)&\longrightarrow(0,\infty)\\
t&\longmapsto\sum_{j=0}^{N-1}w_j e^{-x_jt}, \tag*{(88)}
\end{aligned} k ε , R , η : ( 0 , ∞ ) t ⟶ ( 0 , ∞ ) ⟼ j = 0 ∑ N − 1 w j e − x j t , (88) where
w j : = 2 H Γ ( β + 1 ) ( x j + 1 β − x j β ) > 0. w_j:=\frac{\sqrt{2H}}{\Gamma(\beta+1)}\left(x_{j+1}^{\beta}-x_j^{\beta}\right)>0. w j := Γ ( β + 1 ) 2 H ( x j + 1 β − x j β ) > 0. Let Z ε , R , η Z^{\varepsilon,R,\eta} Z ε , R , η be the target built from this kernel using stationary factors, with initial value zero. There is C H < ∞ C_H<\infty C H < ∞ such that, for L ≥ 1 L\ge1 L ≥ 1 ,
sup 0 ≤ t ≤ L ∥ Z t ε , R , η − B t H ∥ L 2 ≤ C H { L ε β + R − H + η L H } . (89) \sup_{0\le t\le L}\left\|Z_t^{\varepsilon,R,\eta}-B_t^H\right\|_{L^2}\le C_H\left\{\sqrt{L}\,\varepsilon^\beta+R^{-H}+\eta L^H\right\}. \tag*{(89)} 0 ≤ t ≤ L sup Z t ε , R , η − B t H L 2 ≤ C H { L ε β + R − H + η L H } . (89) In particular, fix γ > 0 \gamma>0 γ > 0 and, for L ≥ 2 L\ge2 L ≥ 2 , choose
ε L = L − ( γ + 1 / 2 ) / β , R L = L γ / H , η L = L − ( H + γ ) . (90) \varepsilon_L=L^{-(\gamma+1/2)/\beta},\qquad R_L=L^{\gamma/H},\qquad\eta_L=L^{-(H+\gamma)}. \tag*{(90)} ε L = L − ( γ + 1/2 ) / β , R L = L γ / H , η L = L − ( H + γ ) . (90) A geometric partition then uses N L = O H , γ ( L H + γ log L ) N_L=O_{H,\gamma}(L^{H+\gamma}\log L) N L = O H , γ ( L H + γ log L ) factors with positive weights, and its cell value satisfies
∣ u N L ( L ) L − κ H ∣ ≤ C H , γ { L − γ + L − 1 } . (91) \left|\frac{u_{N_L}(L)}{L}-\kappa_H\right|\le C_{H,\gamma}\{L^{-\gamma}+L^{-1}\}. \tag*{(91)} L u N L ( L ) − κ H ≤ C H , γ { L − γ + L − 1 } . (91) Here u N L u_{N_L} u N L denotes the cell value built from this quadrature with N L N_L N L factors. Proof. See § 18 .§ 3 .
Proposition 8.5 (the small-cost limit for fixed factors). Fix n n n , put D n : = A H − 1 ∑ i = 1 N n w i , n > 0 D_n:=A_H^{-1}\sum_{i=1}^{N_n}w_{i,n}>0 D n := A H − 1 ∑ i = 1 N n w i , n > 0 , and define
C n : = m 1 T 2 A H ∑ i = 1 N n w i , n x i , n . C_n:=\frac{m_1T}{\sqrt{2A_H}}\sum_{i=1}^{N_n}w_{i,n}\sqrt{x_{i,n}}. C n := 2 A H m 1 T i = 1 ∑ N n w i , n x i , n . Write V λ ( n ) ( T ) V_\lambda^{(n)}(T) V λ ( n ) ( T ) for the tracking value of Z n Z^n Z n with unit weights and observation filtration F Y n \mathcal{F}^{Y^n} F Y n , without a terminal reset. Write V λ B ( T ; D n ) V_\lambda^B(T;D_n) V λ B ( T ; D n ) for the value of the Brownian target D n W D_nW D n W with unit weights and its natural filtration. Then
∣ V λ ( n ) ( T ) − V λ B ( T ; D n ) ∣ ≤ λ C n . (92) \left|V_\lambda^{(n)}(T)-V_\lambda^B(T;D_n)\right|\le\lambda C_n. \tag*{(92)} V λ ( n ) ( T ) − V λ B ( T ; D n ) ≤ λ C n . (92) Therefore,
V λ ( n ) ( T ) ∼ λ → 0 κ 1 / 2 T D n 4 / 3 λ 2 / 3 . (93) V_\lambda^{(n)}(T)\mathop{\sim}_{\lambda\to0}\kappa_{1/2}TD_n^{4/3}\lambda^{2/3}. \tag*{(93)} V λ ( n ) ( T ) ∼ λ → 0 κ 1/2 T D n 4/3 λ 2/3 . (93) Proof. Integrating the OU equations and using W 0 = 0 W_0=0 W 0 = 0 gives
Y t i , n − Y 0 i , n = W t − x i , n ∫ 0 t Y s i , n d s . Y_t^{i,n}-Y_0^{i,n}=W_t-x_{i,n}\int_0^tY_s^{i,n}\,\mathrm{d}s. Y t i , n − Y 0 i , n = W t − x i , n ∫ 0 t Y s i , n d s . Therefore
Z t n = D n W t + A t n , A t n : = − 1 A H ∑ i = 1 N n w i , n x i , n ∫ 0 t Y s i , n d s . Z_t^n=D_nW_t+A_t^n,\qquad A_t^n:=-\frac{1}{A_H}\sum_{i=1}^{N_n}w_{i,n}x_{i,n}\int_0^tY_s^{i,n}\,\mathrm{d}s. Z t n = D n W t + A t n , A t n := − A H 1 i = 1 ∑ N n w i , n x i , n ∫ 0 t Y s i , n d s . This adapted perturbation is absolutely continuous and starts at zero. Stationarity and the Wiener isometry give
E [ ( Y s i , n ) 2 ] = ∫ 0 ∞ e − 2 x i , n r d r = 1 2 x i , n . \mathbb{E}\left[(Y_s^{i,n})^2\right]=\int_0^\infty e^{-2x_{i,n}r}\,\mathrm{d}r=\frac{1}{2x_{i,n}}. E [ ( Y s i , n ) 2 ] = ∫ 0 ∞ e − 2 x i , n r d r = 2 x i , n 1 . Then,
E [ ∣ Y s i , n ∣ ] = m 1 2 x i , n . \mathbb{E}\left[|Y_s^{i,n}|\right]=\frac{m_1}{\sqrt{2x_{i,n}}}. E [ ∣ Y s i , n ∣ ] = 2 x i , n m 1 .
The triangle inequality and these stationary Gaussian moments now yield
E [ TV ( 0 , T ] ( A n ) ] = E [ ∫ 0 T ∣ 1 A H ∑ i = 1 N n w i , n x i , n Y s i , n ∣ d s ] ≤ 1 A H ∑ i = 1 N n w i , n x i , n ∫ 0 T E [ ∣ Y s i , n ∣ ] d s = m 1 T 2 A H ∑ i = 1 N n w i , n x i , n = C n . \begin{aligned}
\mathbb{E}\left[\operatorname{TV}_{(0,T]}(A^n)\right] = \mathbb{E}\left[\int_0^T \left|\frac{1}{A_H}\sum_{i=1}^{N_n} w_{i,n}x_{i,n}Y_s^{i,n}\right|\,\mathrm{d}s\right] \\
&\le\frac{1}{A_H}\sum_{i=1}^{N_n}w_{i,n}x_{i,n}\int_0^T\mathbb{E}\left[|Y_s^{i,n}|\right]\,\mathrm{d}s \\
&= \frac{m_1T}{\sqrt{2}A_H}\sum_{i=1}^{N_n}w_{i,n}\sqrt{x_{i,n}} = C_n.
\end{aligned} E [ TV ( 0 , T ] ( A n ) ] = E [ ∫ 0 T A H 1 i = 1 ∑ N n w i , n x i , n Y s i , n d s ] ≤ A H 1 i = 1 ∑ N n w i , n x i , n ∫ 0 T E [ ∣ Y s i , n ∣ ] d s = 2 A H m 1 T i = 1 ∑ N n w i , n x i , n = C n .
Apply proposition 2.4 in the full-past filtration. The negative Brownian path is independent of the future increments. Lemma 6.4 removes it from the Brownian value. Thus (92) follows, and
λ − 2 / 3 ∣ V λ ( n ) ( T ) − V λ B ( T ; D n ) ∣ ≤ C n λ 1 / 3 → λ → 0 0. \lambda^{-2/3}\left|V_\lambda^{(n)}(T)-V_\lambda^B(T;D_n)\right|\le C_n\lambda^{1/3}\xrightarrow{\lambda\to0}0. λ − 2/3 V λ ( n ) ( T ) − V λ B ( T ; D n ) ≤ C n λ 1/3 λ → 0 0.
For fixed n n n , theorem 3.4 and proposition 4.1 give the asymptotic formula.
Remark 8.6 (different small-cost exponents). For every fixed n n n , proposition 8.5 gives
lim λ → 0 log V λ ( n ) ( T ) log λ = 2 3 , \lim_{\lambda\to0}\frac{\log V_\lambda^{(n)}(T)}{\log\lambda}=\frac{2}{3}, λ → 0 lim log λ log V λ ( n ) ( T ) = 3 2 ,
whereas theorem 3.4 gives exponent 2 H / ( 1 + H ) 2H/(1+H) 2 H / ( 1 + H ) for fBm. The joint limit in theorem 8.2 couples resolution with horizon growth.
Proposition 8.7 (Markov representation of the factor cell). Let
F t n , + : = σ P ( Y 0 n , W s − W 0 : 0 ≤ s ≤ t ) = σ P ( Y s n : 0 ≤ s ≤ t ) , 0 ≤ t ≤ L . \mathcal{F}_t^{n,+}:=\sigma^{\mathbb{P}}\left(Y_0^n,W_s-W_0:0\le s\le t\right)
=\sigma^{\mathbb{P}}\left(Y_s^n:0\le s\le t\right),\qquad0\le t\le L. F t n , + := σ P ( Y 0 n , W s − W 0 : 0 ≤ s ≤ t ) = σ P ( Y s n : 0 ≤ s ≤ t ) , 0 ≤ t ≤ L . The second equality follows from the factor SDE and the reconstruction formula used in lemma 8.1. Then
U L ( Z n ; F Y n ) = U L ( Z n ; F n , + ) . (94) \mathcal{U}_L(Z^n;\mathbb{F}^{Y^n})=\mathcal{U}_L(Z^n;\mathbb{F}^{n,+}). \tag*{(94)} U L ( Z n ; F Y n ) = U L ( Z n ; F n , + ) . (94) Thus the reset value depends on the past before zero only through Y 0 n Y_0^n Y 0 n .
Proof. The factor solution is
Y t i , n = e − x i , n t Y 0 i , n + ∫ 0 t e − x i , n ( t − s ) d W s , t ≥ 0. Y_t^{i,n}=e^{-x_{i,n}t}Y_0^{i,n}+\int_0^t e^{-x_{i,n}(t-s)}\,\mathrm{d}W_s,\qquad t\ge0. Y t i , n = e − x i , n t Y 0 i , n + ∫ 0 t e − x i , n ( t − s ) d W s , t ≥ 0.
Thus Z n ∣ [ 0 , L ] Z^n|_{[0,L]} Z n ∣ [ 0 , L ] is adapted to F n , + \mathbb{F}^{n,+} F n , + . Let U : = ( Y s n ) s ≤ 0 U:=(Y_s^n)_{s\le0} U := ( Y s n ) s ≤ 0 , viewed in C l o c ( ( − ∞ , 0 ] ; R N n ) \mathcal{C}_{\mathrm{loc}}((-\infty,0];\mathbb{R}^{N_n}) C loc (( − ∞ , 0 ] ; R N n ) . Future Brownian increments are independent of the whole initial past, so
U ⊥ ⊥ F L n , + ∨ σ P ( Y 0 n ) , F t Y n = F t n , + ∨ σ P ( U ) . U\perp\!\!\!\perp\mathcal{F}_L^{n,+}\vee\sigma^{\mathbb{P}}(Y_0^n),\qquad\mathcal{F}_t^{Y^n}=\mathcal{F}_t^{n,+}\vee\sigma^{\mathbb{P}}(U). U ⊥ ⊥ F L n , + ∨ σ P ( Y 0 n ) , F t Y n = F t n , + ∨ σ P ( U ) .
Apply lemma 6.4 with the initial state Y = Y 0 n Y=Y_0^n Y = Y 0 n to obtain (94) .
The controlled Markov state lies in R × R N n \mathbb{R}\times\mathbb{R}^{N_n} R × R N n . Introduce the coefficient vector
a : = ( w 1 , n / A H , … , w N n , n / A H ) ⊤ ∈ ( 0 , ∞ ) N n a:=(w_{1,n}/A_H,\ldots,w_{N_n,n}/A_H)^\top\in(0,\infty)^{N_n} a := ( w 1 , n / A H , … , w N n , n / A H ) ⊤ ∈ ( 0 , ∞ ) N n
and the diagonal matrix
X n : = diag ( x 1 , n , … , x N n , n ) ∈ R N n × N n . X_n:=\operatorname{diag}(x_{1,n},\ldots,x_{N_n,n})\in\mathbb{R}^{N_n\times N_n}. X n := diag ( x 1 , n , … , x N n , n ) ∈ R N n × N n .
Here 1 ∈ R N n \mathbf{1}\in\mathbb{R}^{N_n} 1 ∈ R N n is the all-ones vector and D n = 1 ⊤ a D_n=\mathbf{1}^\top a D n = 1 ⊤ a . Since Z t n = a ⊤ ( Y t n − Y 0 n ) Z_t^n=a^\top(Y_t^n-Y_0^n) Z t n = a ⊤ ( Y t n − Y 0 n ) , the error e : = ψ − Z n e:=\psi-Z^n e := ψ − Z n satisfies
d e t = d ψ t + a ⊤ X n Y t n d t − D n d W t , (95) \mathrm{d}e_t=\mathrm{d}\psi_t+a^\top X_nY_t^n\,\mathrm{d}t-D_n\,\mathrm{d}W_t, \tag*{(95)} d e t = d ψ t + a ⊤ X n Y t n d t − D n d W t , (95)
and
d Y t n = − X n Y t n d t + 1 d W t . (96) \mathrm{d}Y_t^n=-X_nY_t^n\,\mathrm{d}t+\mathbf{1}\,\mathrm{d}W_t. \tag*{(96)} d Y t n = − X n Y t n d t + 1 d W t . (96)
The initial vector Y 0 n Y_0^n Y 0 n is centered Gaussian, and the Wiener isometry gives its covariance:
E [ Y 0 i , n Y 0 j , n ] = ∫ − ∞ 0 e ( x i , n + x j , n ) s d s = 1 x i , n + x j , n . \mathbb{E}\left[Y_0^{i,n}Y_0^{j,n}\right]=\int_{-\infty}^{0}e^{(x_{i,n}+x_{j,n})s}\,\mathrm{d}s=\frac{1}{x_{i,n}+x_{j,n}}. E [ Y 0 i , n Y 0 j , n ] = ∫ − ∞ 0 e ( x i , n + x j , n ) s d s = x i , n + x j , n 1 .
For f ∈ C 2 ( R N n + 1 ; R ) f\in\mathcal{C}^2(\mathbb{R}^{N_n+1};\mathbb{R}) f ∈ C 2 ( R N n + 1 ; R ) , the generator between interventions is
L n f ( e , y ) = ( a ⊤ X n y ) ∂ e f − ( X n y ) ⋅ ∇ y f + 1 2 ( − D n ∂ e + ∑ i = 1 N n ∂ y i ) 2 f = ( a ⊤ X n y ) ∂ e f − ( X n y ) ⋅ ∇ y f + D n 2 2 ∂ e e f − D n ∑ i = 1 N n ∂ e y i f + 1 2 ∑ i = 1 N n ∑ j = 1 N n ∂ y i y j f . (97) \begin{aligned}
\mathcal{L}_n f(e,y)&=(a^\top X_ny)\partial_e f-(X_ny)\cdot\nabla_y f+\frac{1}{2}\left(-D_n\partial_e+\sum_{i=1}^{N_n}\partial_{y_i}\right)^2f\\
&=(a^\top X_ny)\partial_e f-(X_ny)\cdot\nabla_y f+\frac{D_n^2}{2}\partial_{ee}f-D_n\sum_{i=1}^{N_n}\partial_{ey_i}f+\frac{1}{2}\sum_{i=1}^{N_n}\sum_{j=1}^{N_n}\partial_{y_iy_j}f. \tag*{(97)}
\end{aligned} L n f ( e , y ) = ( a ⊤ X n y ) ∂ e f − ( X n y ) ⋅ ∇ y f + 2 1 ( − D n ∂ e + i = 1 ∑ N n ∂ y i ) 2 f = ( a ⊤ X n y ) ∂ e f − ( X n y ) ⋅ ∇ y f + 2 D n 2 ∂ ee f − D n i = 1 ∑ N n ∂ e y i f + 2 1 i = 1 ∑ N n j = 1 ∑ N n ∂ y i y j f . (97)
The running, variation, and terminal costs are e 2 / 2 e^2/2 e 2 /2 , one per unit adjustment, and ∣ e ∣ |e| ∣ e ∣ .
A quadratic-hedging application. Under the assumptions of section J, the Galtchouk–Kunita–Watanabe decomposition identifies the penalized mean-square hedging value, after subtracting the irreducible term, with V λ ( Θ ) V_\lambda(\Theta) V λ ( Θ ) . The identity and its small-cost consequence are given in section J.
Conclusion
The increment scale determines the exponent, while a causal fractional cell determines the constant. We use resets to obtain subadditivity and finite-memory approximation to show that the initial past has no long-run effect on the value. Averaging over shifted partitions then yields a bounded full-past correction and continuity in H H H . Coupling local paths with their filtrations lets us transfer the cell value to nonstationary targets. Homogeneity then determines the near-optimal cost split. Finally, quadrature with positive weights approximates the cell when resolution grows with the horizon, whereas fixed factors retain a Brownian small-cost limit.
Causal deconvolution of the Mandelbrot–Van Ness map
Lemma A.1 (Brownian and fractional pasts generate the same sigma-field). Fix H ∈ ( 0 , 1 ) H\in(0,1) H ∈ ( 0 , 1 ) . Let W = ( W t ) t ∈ R W=(W_t)_{t\in\mathbb{R}} W = ( W t ) t ∈ R be a two-sided standard Brownian motion with W 0 = 0 W_0=0 W 0 = 0 , and define
B t H : = 1 H ∫ R g t H ( s ) d W s , g t H ( s ) : = ( t − s ) + H − 1 / 2 − ( − s ) + H − 1 / 2 , B_t^H:=\frac{1}{\sqrt{H}}\int_{\mathbb{R}}g_t^H(s)\,\mathrm{d}W_s,\qquad g_t^H(s):=(t-s)_+^{H-1/2}-(-s)_+^{H-1/2}, B t H := H 1 ∫ R g t H ( s ) d W s , g t H ( s ) := ( t − s ) + H − 1/2 − ( − s ) + H − 1/2 , with the order-zero convention from (64) . For t ∈ [ 0 , ∞ ) t\in[0,\infty) t ∈ [ 0 , ∞ ) , set
H t W : = span ‾ L 2 ( Ω ) { W s : s ≤ t } , H t B : = span ‾ L 2 ( Ω ) { B s H : s ≤ t } . \mathcal{H}_t^W:=\overline{\operatorname{span}}^{L^2(\Omega)}\{W_s:s\leq t\},\qquad\mathcal{H}_t^B:=\overline{\operatorname{span}}^{L^2(\Omega)}\{B_s^H:s\leq t\}. H t W := span L 2 ( Ω ) { W s : s ≤ t } , H t B := span L 2 ( Ω ) { B s H : s ≤ t } . The Gaussian spans and completed pasts coincide:
H t W = H t B , σ ( W s : s ≤ t ) ‾ P = σ ( B s H : s ≤ t ) ‾ P . (98) \mathcal{H}_t^W=\mathcal{H}_t^B,\qquad\overline{\sigma(W_s:s\leq t)}^{\mathbb{P}}=\overline{\sigma(B_s^H:s\leq t)}^{\mathbb{P}}. \tag*{(98)} H t W = H t B , σ ( W s : s ≤ t ) P = σ ( B s H : s ≤ t ) P . (98) The result also holds after shifting the time origin, subtracting the value at that origin, and scaling time, for nonnegative relative control times.
Proof. To compare the two pasts, put β : = H − 1 / 2 \beta:=H-1/2 β := H − 1/2 and introduce the oriented Brownian interval indicator , for s ∈ R s\in\mathbb{R} s ∈ R , by
ι s ( r ) : = { 1 ( 0 , s ] ( r ) , s ≥ 0 , − 1 ( s , 0 ] ( r ) , s < 0. \iota_s(r):=
\begin{cases}
\mathbf{1}_{(0,s]}(r), & s\geq0,\\
-\mathbf{1}_{(s,0]}(r), & s<0.
\end{cases} ι s ( r ) := { 1 ( 0 , s ] ( r ) , − 1 ( s , 0 ] ( r ) , s ≥ 0 , s < 0.
Then,
W s = ∫ R ι s ( r ) d W r . W_s=\int_{\mathbb{R}}\iota_s(r)\,\mathrm{d}W_r. W s = ∫ R ι s ( r ) d W r .
For t ≥ 0 t\geq0 t ≥ 0 , these indicators with s ≤ t s\leq t s ≤ t span a dense subspace of L 2 ( ( − ∞ , t ] ) L^2((-\infty,t]) L 2 (( − ∞ , t ]) , so the Wiener isometry gives
H t W = { ∫ R f ( r ) d W r : f ∈ L 2 ( R ) , f = 0 a.e. on ( t , ∞ ) } . \mathcal{H}_t^W=\left\{\int_{\mathbb{R}}f(r)\,\mathrm{d}W_r:f\in L^2(\mathbb{R}),\ f=0\ \text{a.e. on }(t,\infty)\right\}. H t W = { ∫ R f ( r ) d W r : f ∈ L 2 ( R ) , f = 0 a.e. on ( t , ∞ ) } .
Since g s H g_s^H g s H vanishes on ( s ∨ 0 , ∞ ) (s \vee0,\infty) ( s ∨ 0 , ∞ ) , B s H ∈ H t W B_s^H \in\mathcal{H}_t^W B s H ∈ H t W for every s ≤ t s \leq t s ≤ t , hence H t B ⊂ H t W \mathcal{H}_t^B \subset\mathcal{H}_t^W H t B ⊂ H t W . For the converse, we work in the deterministic Wiener-integrand space. Starting from a step function f = ∑ j = 0 M − 1 a j 1 ( r j , r j + 1 ] f=\sum_{j=0}^{M-1}a_j1_{(r_j,r_{j+1}]} f = ∑ j = 0 M − 1 a j 1 ( r j , r j + 1 ] , where M ∈ N M\in\mathbb{N} M ∈ N and r 0 < ⋯ < r M r_0<\cdots<r_M r 0 < ⋯ < r M , define
I B ( f ) : = ∑ j = 0 M − 1 a j ( B r j + 1 H − B r j H ) , ∥ f ∥ H H 2 : = E [ I B ( f ) 2 ] . I_B(f):=\sum_{j=0}^{M-1}a_j(B_{r_{j+1}}^H-B_{r_j}^H),\qquad\lVert f\rVert_{\mathcal{H}^H}^2:=\mathbb{E}[I_B(f)^2]. I B ( f ) := j = 0 ∑ M − 1 a j ( B r j + 1 H − B r j H ) , ∥ f ∥ H H 2 := E [ I B ( f ) 2 ] .
Let H ~ \widetilde{\mathcal{H}} H be the Hilbert completion after identifying functions of zero seminorm. The map I B : H ~ → L 2 ( Ω ) I_B:\widetilde{\mathcal{H}}\to L^2(\Omega) I B : H → L 2 ( Ω ) is its isometric extension. The Wiener-integral deconvolution identity of [34 ] , in the present normalization, is
W s = Γ H Γ ( 1 + β ) Γ ( 1 − β ) ∫ R { ( s − r ) + β − ( − r ) + β } d B r H . (99) W_s=\frac{\sqrt{\Gamma_H}}{\Gamma(1+\beta)\Gamma(1-\beta)}\int_{\mathbb{R}}\{(s-r)_+^\beta-(-r)_+^\beta\}\,\mathrm{d}B_r^H . \tag*{(99)} W s = Γ ( 1 + β ) Γ ( 1 − β ) Γ H ∫ R {( s − r ) + β − ( − r ) + β } d B r H . (99)
This is an L 2 L^2 L 2 Wiener integral. The inverse kernel is supported on ( − ∞ , s ∨ 0 ] (-\infty,s\vee0] ( − ∞ , s ∨ 0 ] , and its step approximations f m f_m f m use only that half-line. By the Wiener isometry,
∥ W s − Γ H I B ( f m ) Γ ( 1 + β ) Γ ( 1 − β ) ∥ L 2 2 = Γ H Γ ( 1 + β ) 2 Γ ( 1 − β ) 2 ∥ ( s − ⋅ ) + β − ( − ⋅ ) + β − f m ∥ H ~ 2 → m → ∞ 0. \left\lVert W_s-\frac{\sqrt{\Gamma_H}I_B(f_m)}{\Gamma(1+\beta)\Gamma(1-\beta)}\right\rVert_{L^2}^2
=\frac{\Gamma_H}{\Gamma(1+\beta)^2\Gamma(1-\beta)^2}\left\lVert(s-\cdot)_+^\beta-(-\cdot)_+^\beta-f_m\right\rVert_{\widetilde{\mathcal{H}}}^2 \xrightarrow[m\to\infty]{}0. W s − Γ ( 1 + β ) Γ ( 1 − β ) Γ H I B ( f m ) L 2 2 = Γ ( 1 + β ) 2 Γ ( 1 − β ) 2 Γ H ( s − ⋅ ) + β − ( − ⋅ ) + β − f m H 2 m → ∞ 0.
For s ≤ t s\leq t s ≤ t and t ≥ 0 t\geq0 t ≥ 0 , each I B ( f m ) I_B(f_m) I B ( f m ) is a finite linear combination of fractional observations at times at most t t t . Thus W s ∈ H t B W_s\in\mathcal{H}_t^B W s ∈ H t B . The normalization in (99) comes from composing the right-sided fractional operators of orders β \beta β and − β -\beta − β , whose kernel factors are Γ ( 1 + β ) \Gamma(1+\beta) Γ ( 1 + β ) and Γ ( 1 − β ) \Gamma(1-\beta) Γ ( 1 − β ) . At β = 0 \beta=0 β = 0 , Γ 1 / 2 = 1 \Gamma_{1/2}=1 Γ 1/2 = 1 and the identity reduces to the Brownian interval integral. The two inclusions show that the closed Gaussian spans coincide. Since an L 2 L^2 L 2 limit has an almost-surely convergent subsequence, completing a Gaussian linear span does not change its completed generated sigma-field. This proves both equalities in (98) . The same support argument applies after shifting the time origin, subtracting the value at that origin, and rescaling, for nonnegative relative times.
Conditional removal of auxiliary variables
Proof of lemma 6.4. Conditional averaging. Let ψ ∈ A L ( G U ) \psi\in\mathcal{A}_L(\mathcal{G}^U) ψ ∈ A L ( G U ) have finite cost. To average out the auxiliary variable, represent the countable base generators by a standard Borel variable Z Z Z and choose a Borel path map ψ = Ψ ( Z , U ) \psi=\Psi(Z,U) ψ = Ψ ( Z , U ) . Conditional independence then gives L ( U ∣ G L ) = η Y \mathcal{L}(U\mid\mathcal{G}_L)=\eta_Y L ( U ∣ G L ) = η Y , where η y : = L ( U ∣ Y = y ) \eta_y:=\mathcal{L}(U\mid Y=y) η y := L ( U ∣ Y = y ) is a regular conditional distribution [32 ] . We therefore define
ψ ˉ t : = ∫ S Ψ ( Z , u ) t η Y ( d u ) . \bar{\psi}_t:=\int_S\Psi(Z,u)_t\,\eta_Y(\mathrm{d}u). ψ ˉ t := ∫ S Ψ ( Z , u ) t η Y ( d u ) .
As ψ 0 − = 0 \psi_{0-}=0 ψ 0 − = 0 ,
sup t ≤ L ∣ ψ t ∣ ≤ TV [ 0 , L ] a u g ( ψ ) , and E [ TV [ 0 , L ] a u g ( ψ ) ] < ∞ . \sup_{t\leq L}|\psi_t|\leq\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi),\qquad\text{and}\qquad\mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)\right]<\infty. t ≤ L sup ∣ ψ t ∣ ≤ TV [ 0 , L ] aug ( ψ ) , and E [ TV [ 0 , L ] aug ( ψ ) ] < ∞.
Conditional Fubini supplies one integrable pathwise dominating function for the average. Dominated convergence gives right-continuity and left limits of ψ ˉ \bar{\psi} ψ ˉ . For every finite partition π \pi π of [ 0 , L ] [0,L] [ 0 , L ] whose interior points are rational,
∣ ψ ˉ 0 ∣ + ∑ [ s , t ] ∈ π ∣ ψ ˉ t − ψ ˉ s ∣ ≤ ∫ S ( ∣ Ψ ( Z , u ) 0 ∣ + ∑ [ s , t ] ∈ π ∣ Ψ ( Z , u ) t − Ψ ( Z , u ) s ∣ ) η Y ( d u ) . |\bar{\psi}_0|+\sum_{[s,t]\in\pi}|\bar{\psi}_t-\bar{\psi}_s|
\leq\int_S\left(|\Psi(Z,u)_0|+\sum_{[s,t]\in\pi}|\Psi(Z,u)_t-\Psi(Z,u)_s|\right)\eta_Y(\mathrm{d}u). ∣ ψ ˉ 0 ∣ + [ s , t ] ∈ π ∑ ∣ ψ ˉ t − ψ ˉ s ∣ ≤ ∫ S ∣Ψ ( Z , u ) 0 ∣ + [ s , t ] ∈ π ∑ ∣Ψ ( Z , u ) t − Ψ ( Z , u ) s ∣ η Y ( d u ) .
Taking the supremum over these partitions yields
TV [ 0 , L ] a u g ( ψ ˉ ) ≤ ∫ S TV [ 0 , L ] a u g ( Ψ ( Z , u ) ) η Y ( d u ) . \operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\bar{\psi})\leq\int_S\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\Psi(Z,u))\eta_Y(\mathrm{d}u). TV [ 0 , L ] aug ( ψ ˉ ) ≤ ∫ S TV [ 0 , L ] aug ( Ψ ( Z , u )) η Y ( d u ) .
The same argument on subintervals gives domination of the variation measures and hence of their integrals against deterministic nonnegative Borel weights. Set Q L : = ( Q ∩ [ 0 , L ) ) ∪ { L } Q_L:=(\mathbb{Q}\cap[0,L))\cup\{L\} Q L := ( Q ∩ [ 0 , L )) ∪ { L } . For q ∈ Q L q\in Q_L q ∈ Q L , adaptedness represents Ψ ( Z , U ) q \Psi(Z,U)_q Ψ ( Z , U ) q using only G q \mathcal{G}_q G q and U U U . Since Y Y Y is G 0 \mathcal{G}_0 G 0 -measurable, integration against η Y \eta_Y η Y makes ψ ˉ q \bar{\psi}_q ψ ˉ q G q \mathcal{G}_q G q -measurable. A common null set suffices for all q ∈ Q L q\in Q_L q ∈ Q L . Right-continuity of paths and filtration extends adaptedness to every t < L t<L t < L , while the terminal time is included in Q L Q_L Q L . Jensen also gives
E [ ∫ 0 L ψ ˉ t 2 d t ] ≤ E [ ∫ 0 L ψ t 2 d t ] , \mathbb{E}\left[\int_0^L\bar{\psi}_t^2\,\mathrm{d}t\right]\leq\mathbb{E}\left[\int_0^L\psi_t^2\,\mathrm{d}t\right], E [ ∫ 0 L ψ ˉ t 2 d t ] ≤ E [ ∫ 0 L ψ t 2 d t ] ,
and
E [ ∣ ψ ˉ L ∣ ] ≤ E [ ∣ ψ L ∣ ] . \mathbb{E}[|\bar{\psi}_L|]\leq\mathbb{E}[|\psi_L|]. E [ ∣ ψ ˉ L ∣ ] ≤ E [ ∣ ψ L ∣ ] .
Thus ψ ˉ ∈ A L ( G ) \bar{\psi}\in\mathcal{A}_L(\mathcal{G}) ψ ˉ ∈ A L ( G ) . The target is G L \mathcal{G}_L G L -measurable. Conditional Jensen for the quadratic and absolute-value terms, together with the variation bound, yields
C L X ( ψ ˉ ) ≤ ∫ S C L X ( Ψ ( Z , u ) ) η Y ( d u ) = E [ C L X ( ψ ) ∣ G L ] . \mathcal{C}_L^X(\bar{\psi})\leq\int_S\mathcal{C}_L^X(\Psi(Z,u))\eta_Y(\mathrm{d}u)=\mathbb{E}[\mathcal{C}_L^X(\psi)\mid\mathcal{G}_L]. C L X ( ψ ˉ ) ≤ ∫ S C L X ( Ψ ( Z , u )) η Y ( d u ) = E [ C L X ( ψ ) ∣ G L ] .
Taking expectations proves the reverse inequality in (46) . Conditioning on Y Y Y proves the reverse inequality in (47) . The other inequalities follow from inclusion of the control classes. Initial conditional values. Let V V V be the base conditional value given G 0 \mathcal{G}_0 G 0 , and W W W the enlarged conditional value given G 0 U \mathcal{G}^{U}_0 G 0 U . For integrable G L \mathcal{G}_L G L -measurable C C C ,
E [ C ∣ G 0 U ] = E [ C ∣ G 0 ] , \mathbb{E}[C \mid\mathcal{G}^{U}_0] = \mathbb{E}[C \mid\mathcal{G}_0], E [ C ∣ G 0 U ] = E [ C ∣ G 0 ] ,
so testing enlarged values with base controls gives W ≤ V W \le V W ≤ V . The conditional cost families are downward directed: paste two controls on the event, known initially, where the first conditional cost is smaller. They therefore admit decreasing sequences approaching their essential infima, bounded above by the conditional cost of a fixed finite-cost control. Dominated convergence then gives
E [ W ] = U L ( X ; G U ) = U L ( X ; G ) = E [ V ] . \mathbb{E}[W] = \mathcal{U}_L(X;\mathcal{G}^{U}) = \mathcal{U}_L(X;\mathcal{G}) = \mathbb{E}[V]. E [ W ] = U L ( X ; G U ) = U L ( X ; G ) = E [ V ] .
Since 0 ≤ W ≤ V 0 \le W \le V 0 ≤ W ≤ V and both are integrable, W = V W = V W = V almost surely. For a G 0 \mathcal{G}_0 G 0 -measurable pre-initial position x 0 x_0 x 0 , average ψ − x 0 \psi-x_0 ψ − x 0 and add x 0 x_0 x 0 back. The same convexity arguments apply to the weighted costs whenever a finite-cost admissible control exists.
Path-space rules and concatenation
Fix L > 0 L>0 L > 0 . As above, write C L = C ( [ 0 , L ] ; R ) \mathcal{C}_L=\mathcal{C}([0,L];\mathbb{R}) C L = C ([ 0 , L ] ; R ) and D L = D ( [ 0 , L ] ; R ) \mathcal{D}_L=\mathcal{D}([0,L];\mathbb{R}) D L = D ([ 0 , L ] ; R ) , and put Q L : = ( Q ∩ [ 0 , L ) ) ∪ { L } Q_L:=(Q\cap[0,L))\cup\{L\} Q L := ( Q ∩ [ 0 , L )) ∪ { L } . These path spaces are Polish under the uniform and Skorokhod J 1 J_1 J 1 topologies, respectively [35 Chapters 2–3] . For μ ∈ P ( C L ) \mu\in\mathcal{P}(\mathcal{C}_L) μ ∈ P ( C L ) , introduce the coordinate maps
X t : C L ⟶ R x ⟼ x ( t ) , 0 ≤ t ≤ L . \begin{aligned}
X_t:\quad&\mathcal{C}_L \longrightarrow\mathbb{R}\\
&x \longmapsto x(t), \qquad0\le t\le L.
\end{aligned} X t : C L ⟶ R x ⟼ x ( t ) , 0 ≤ t ≤ L .
The coordinate maps then induce the raw and completed filtrations
G t 0 : = σ ( X s : 0 ≤ s ≤ t ) , G t μ : = G t 0 ∨ N μ , \mathcal{G}^{0}_t:=\sigma(X_s:0\le s\le t), \qquad\mathcal{G}^{\mu}_t:=\mathcal{G}^{0}_t\vee\mathcal{N}^{\mu}, G t 0 := σ ( X s : 0 ≤ s ≤ t ) , G t μ := G t 0 ∨ N μ ,
where
N μ : = { N ⊂ C L : ∃ N 0 ∈ G L 0 , N ⊂ N 0 , μ ( N 0 ) = 0 } . \mathcal{N}^{\mu}:=\{N\subset\mathcal{C}_L:\exists N_0\in\mathcal{G}^{0}_L,\ N\subset N_0,\ \mu(N_0)=0\}. N μ := { N ⊂ C L : ∃ N 0 ∈ G L 0 , N ⊂ N 0 , μ ( N 0 ) = 0 } .
We assume that ( G t μ ) 0 ≤ t ≤ L (\mathcal{G}^{\mu}_t)_{0\le t\le L} ( G t μ ) 0 ≤ t ≤ L is right-continuous. For fBm this follows from its causal Volterra representation and inverse [31 , 34 ] . On D L \mathcal{D}_L D L , the coordinate evaluations at times in Q L Q_L Q L generate B ( D L ) \mathcal{B}(\mathcal{D}_L) B ( D L ) [35 Chapter 3] . Let Π ( Q L ) \Pi(Q_L) Π ( Q L ) be the countable set of finite partitions 0 = q 0 < ⋯ < q n = L 0=q_0<\cdots<q_n=L 0 = q 0 < ⋯ < q n = L with q j ∈ Q L q_j\in Q_L q j ∈ Q L . Right-continuity gives
TV ( 0 , L ] ( z ) = sup π ∈ Π ( Q L ) ∑ j = 1 n ( π ) ∣ z q j − z q j − 1 ∣ . \operatorname{TV}_{(0,L]}(z)=\sup_{\pi\in\Pi(Q_L)}\sum_{j=1}^{n(\pi)}\lvert z_{q_j}-z_{q_{j-1}}\rvert. TV ( 0 , L ] ( z ) = π ∈ Π ( Q L ) sup j = 1 ∑ n ( π ) ∣ z q j − z q j − 1 ∣ .
Because the supremum is countable, the variation is Borel. Consequently,
{ z ∈ D L : TV ( 0 , L ] ( z ) < ∞ } = ⋃ m ∈ N { z : TV ( 0 , L ] ( z ) ≤ m } ∈ B ( D L ) . \{z\in\mathcal{D}_L:\operatorname{TV}_{(0,L]}(z)<\infty\}
=\bigcup_{m\in\mathbb{N}}\{z:\operatorname{TV}_{(0,L]}(z)\le m\}\in\mathcal{B}(\mathcal{D}_L). { z ∈ D L : TV ( 0 , L ] ( z ) < ∞ } = m ∈ N ⋃ { z : TV ( 0 , L ] ( z ) ≤ m } ∈ B ( D L ) .
Theorem C.1 (path-space representation and law transport). Let X X X be the canonical coordinate just defined and let ψ = ( ψ t ) 0 ≤ t ≤ L \psi=(\psi_t)_{0\le t\le L} ψ = ( ψ t ) 0 ≤ t ≤ L be a process on C L \mathcal{C}_L C L that is adapted to ( G t μ ) (\mathcal{G}^{\mu}_t) ( G t μ ) , right-continuous, of finite variation μ \mu μ -a.s., and integrable for the relevant cost. There is a Borel map Φ : C L → D L \Phi:\mathcal{C}_L\to\mathcal{D}_L Φ : C L → D L such that
μ ( Φ ( X ) t = ψ t for all t ∈ [ 0 , L ] ) = 1. \mu(\Phi(X)_t=\psi_t\text{ for all }t\in[0,L])=1. μ ( Φ ( X ) t = ψ t for all t ∈ [ 0 , L ]) = 1. There is a μ \mu μ -null set outside which the map is causal at every q ∈ Q L q\in Q_L q ∈ Q L : for each q ∈ Q L q\in Q_L q ∈ Q L , there is a Borel map f q : C ( [ 0 , q ] ; R ) → R f_q:\mathcal{C}([0,q];\mathbb{R})\to\mathbb{R} f q : C ([ 0 , q ] ; R ) → R satisfying
Φ ( X ) q = f q ( X ∣ [ 0 , q ] ) μ -a.s. (100) \Phi(X)_q=f_q(X|_{[0,q]}) \qquad\mu\text{-a.s.} \tag*{(100)} Φ ( X ) q = f q ( X ∣ [ 0 , q ] ) μ -a.s. (100) If Y Y Y is any continuous process with law μ \mu μ , defined on any complete probability space and equipped with its usual natural filtration, then Φ ( Y ) \Phi(Y) Φ ( Y ) is adapted and
( Y , Φ ( Y ) ) = d ( X , Φ ( X ) ) . (101) (Y,\Phi(Y))\mathrel{\overset{d}{=}}(X,\Phi(X)). \tag*{(101)} ( Y , Φ ( Y )) = d ( X , Φ ( X )) . (101) Conversely, every adapted right-continuous finite-variation control on the natural realization of Y Y Y has such a Borel representative on C L \mathcal{C}_L C L .
Proof. Rational coordinate measurability makes ψ \psi ψ a D L \mathcal{D}_L D L -valued variable measurable for the completion of σ ( X ) \sigma(X) σ ( X ) . Standard Borel factorization [32 ] gives a Borel map Φ : C L → D L \Phi:\mathcal{C}_L\to\mathcal{D}_L Φ : C L → D L with
ψ = Φ ( X ) μ -a.s. as path-valued variables . \psi=\Phi(X)\qquad\mu\text{-a.s. as path-valued variables}. ψ = Φ ( X ) μ -a.s. as path-valued variables .
For q ∈ Q L q \in Q_L q ∈ Q L , adaptedness and Doob–Dynkin give a Borel map f q : C ( [ 0 , q ] ; R ) → R f_q:\mathcal{C}([0,q];\mathbb{R}) \to\mathbb{R} f q : C ([ 0 , q ] ; R ) → R such that
Φ ( X ) q = f q ( X ∣ [ 0 , q ] ) μ -a.s. \Phi(X)_q=f_q(X|_{[0,q]})\quad\mu\text{-a.s.} Φ ( X ) q = f q ( X ∣ [ 0 , q ] ) μ -a.s.
We can combine the exceptional sets into one null set because Q L Q_L Q L is countable. Equality in law transfers all these identities to Y Y Y . For t < L t<L t < L , choose q n ↓ t q_n \downarrow t q n ↓ t in Q L Q_L Q L . Then
Φ ( Y ) t = lim n → ∞ Φ ( Y ) q n , ⋂ n = 1 ∞ F q n Y = F t Y . \Phi(Y)_t=\lim_{n\to\infty}\Phi(Y)_{q_n},\qquad\bigcap_{n=1}^{\infty}\mathcal{F}^{Y}_{q_n}=\mathcal{F}^{Y}_t. Φ ( Y ) t = n → ∞ lim Φ ( Y ) q n , n = 1 ⋂ ∞ F q n Y = F t Y .
The limit is measurable for each F q n Y \mathcal{F}^{Y}_{q_n} F q n Y , hence for F t Y \mathcal{F}^{Y}_t F t Y . Thus Φ ( Y ) \Phi(Y) Φ ( Y ) is adapted. The terminal time is covered because L ∈ Q L L\in Q_L L ∈ Q L . The map x ↦ ( x , Φ ( x ) ) x\mapsto(x,\Phi(x)) x ↦ ( x , Φ ( x )) from C L \mathcal{C}_L C L to C L × D L \mathcal{C}_L\times\mathcal{D}_L C L × D L is Borel. For any Borel A ⊂ C L × D L A\subset\mathcal{C}_L\times\mathcal{D}_L A ⊂ C L × D L ,
P ( ( Y , Φ ( Y ) ) ∈ A ) = ∫ C L 1 A ( x , Φ ( x ) ) μ ( d x ) = μ ( ( X , Φ ( X ) ) ∈ A ) . \mathbb{P}((Y,\Phi(Y))\in A)=\int_{\mathcal{C}_L}\mathbf{1}_A(x,\Phi(x))\mu(dx)=\mu((X,\Phi(X))\in A). P (( Y , Φ ( Y )) ∈ A ) = ∫ C L 1 A ( x , Φ ( x )) μ ( d x ) = μ (( X , Φ ( X )) ∈ A ) .
This proves (100) and preserves admissibility and costs. Conversely, on a natural realization of Y Y Y , factor any admissible D L \mathcal{D}_L D L -valued control through Y Y Y and repeat the rational-time argument on the canonical space. □
Corollary C.2 (equality of natural-filtration values). Let Y Y Y be a continuous process with law μ \mu μ , equipped with its usual natural filtration. Let J : C L × D L → [ 0 , ∞ ] J:\mathcal{C}_L\times\mathcal{D}_L\to[0,\infty] J : C L × D L → [ 0 , ∞ ] be Borel. Then the infimum of E [ J ( Y , ψ ) ] \mathbb{E}[J(Y,\psi)] E [ J ( Y , ψ )] over adapted right-continuous finite-variation controls equals the corresponding infimum on the path space.
Proof. By theorem C.1, every canonical rule transports to an admissible rule for Y Y Y , and every admissible rule for Y Y Y has a canonical representative. In either direction,
E [ J ( Y , Φ ( Y ) ) ] = ∫ C L J ( x , Φ ( x ) ) μ ( d x ) = E μ [ J ( X , Φ ( X ) ) ] . \mathbb{E}[J(Y,\Phi(Y))]=\int_{\mathcal{C}_L}J(x,\Phi(x))\mu(dx)=\mathbb{E}_{\mu}[J(X,\Phi(X))]. E [ J ( Y , Φ ( Y ))] = ∫ C L J ( x , Φ ( x )) μ ( d x ) = E μ [ J ( X , Φ ( X ))] .
Taking infima gives equality. □
Lemma C.3 (path-space concatenation with reset). For S , T > 0 S,T>0 S , T > 0 , let Φ S : C ( [ 0 , S ] ; R ) → D ( [ 0 , S ] ; R ) \Phi^S:\mathcal{C}([0,S];\mathbb{R})\to\mathcal{D}([0,S];\mathbb{R}) Φ S : C ([ 0 , S ] ; R ) → D ([ 0 , S ] ; R ) and Φ T : C ( [ 0 , T ] ; R ) → D ( [ 0 , T ] ; R ) \Phi^T:\mathcal{C}([0,T];\mathbb{R})\to\mathcal{D}([0,T];\mathbb{R}) Φ T : C ([ 0 , T ] ; R ) → D ([ 0 , T ] ; R ) be Borel representatives given by theorem C.1. For x ∈ C S + T x\in\mathcal{C}_{S+T} x ∈ C S + T , write x S : = x ∣ [ 0 , S ] x^S:=x|_{[0,S]} x S := x ∣ [ 0 , S ] and y t : = x S + t − x S y_t:=x_{S+t}-x_S y t := x S + t − x S . Define
Ψ : { C S + T ⟶ D S + T x ⟼ ( t ↦ { Φ S ( x S ) t , 0 ≤ t < S , x S + Φ T ( y ) t − S , S ≤ t ≤ S + T ) . \Psi:\begin{cases}
\mathcal{C}_{S+T}\longrightarrow\mathcal{D}_{S+T}\\
x\longmapsto\left(t\mapsto
\begin{cases}
\Phi^S(x^S)_t,&0\le t<S,\\
x_S+\Phi^T(y)_{t-S},&S\le t\le S+T
\end{cases}\right).
\end{cases} Ψ : ⎩ ⎨ ⎧ C S + T ⟶ D S + T x ⟼ ( t ↦ { Φ S ( x S ) t , x S + Φ T ( y ) t − S , 0 ≤ t < S , S ≤ t ≤ S + T ) . The resulting Ψ \Psi Ψ is Borel and adapted under fBm law. Its running cost splits across the blocks, while its variation satisfies
TV [ 0 , S + T ] a u g ( Ψ ) ≤ TV [ 0 , S ] a u g ( Φ S ( x S ) ) + ∣ Φ S ( x S ) S − x S ∣ + TV [ 0 , T ] a u g ( Φ T ( y ) ) , (102) \operatorname{TV}^{\mathrm{aug}}_{[0,S+T]}(\Psi)\leq\operatorname{TV}^{\mathrm{aug}}_{[0,S]}(\Phi^S(x^S))+|\Phi^S(x^S)_S-x_S|+\operatorname{TV}^{\mathrm{aug}}_{[0,T]}(\Phi^T(y)), \tag*{(102)} TV [ 0 , S + T ] aug ( Ψ ) ≤ TV [ 0 , S ] aug ( Φ S ( x S )) + ∣ Φ S ( x S ) S − x S ∣ + TV [ 0 , T ] aug ( Φ T ( y )) , (102) Moreover, the second rule alone determines the terminal error:
∣ Ψ S + T − x S + T ∣ = ∣ Φ T ( y ) T − y T ∣ . (103) |\Psi_{S+T}-x_{S+T}|=|\Phi^T(y)_T-y_T|. \tag*{(103)} ∣ Ψ S + T − x S + T ∣ = ∣ Φ T ( y ) T − y T ∣. (103)
Proof. Restriction and shifted-increment maps are continuous on the input spaces. Every rational evaluation of Ψ \Psi Ψ is Borel, and the concatenation is càdlàg, so Ψ \Psi Ψ is a Borel D S + T \mathcal{D}_{S+T} D S + T -valued map. Under fBm law, Y r : = X S + r − X S Y_r:=X_{S+r}-X_S Y r := X S + r − X S has law μ H T \mu_H^T μ H T and
σ ( Y v : 0 ≤ v ≤ r ) ⊂ F S + r B H . \sigma(Y_v:0\leq v\leq r)\subset\mathcal{F}^{B_H}_{S+r}. σ ( Y v : 0 ≤ v ≤ r ) ⊂ F S + r B H .
Law transport therefore makes the second rule adapted. At the interface, the triangle inequality gives
∣ x S + Φ T ( y ) 0 − Φ S ( x S ) S − ∣ ≤ ∣ Φ S ( x S ) S − Φ S ( x S ) S − ∣ + ∣ Φ S ( x S ) S − x S ∣ + ∣ Φ T ( y ) 0 ∣ . |x_S+\Phi^T(y)_0-\Phi^S(x^S)_{S-}|\leq|\Phi^S(x^S)_S-\Phi^S(x^S)_{S-}|+|\Phi^S(x^S)_S-x_S|+|\Phi^T(y)_0|. ∣ x S + Φ T ( y ) 0 − Φ S ( x S ) S − ∣ ≤ ∣ Φ S ( x S ) S − Φ S ( x S ) S − ∣ + ∣ Φ S ( x S ) S − x S ∣ + ∣ Φ T ( y ) 0 ∣.
The first term belongs to the first variation, the second is its reset, and the third is the second initial atom. All other increments lie in disjoint blocks, proving (101). To split the running cost, observe that
Ψ t − x t = Φ S ( x S ) t − x t ( 0 ≤ t < S ) , \Psi_t-x_t=\Phi^S(x^S)_t-x_t\qquad(0\leq t<S), Ψ t − x t = Φ S ( x S ) t − x t ( 0 ≤ t < S ) ,
and
Ψ S + r − x S + r = Φ T ( y ) r − y r ( 0 ≤ r ≤ T ) . \Psi_{S+r}-x_{S+r}=\Phi^T(y)_r-y_r\qquad(0\leq r\leq T). Ψ S + r − x S + r = Φ T ( y ) r − y r ( 0 ≤ r ≤ T ) .
These identities split the running integral. Taking r = T r=T r = T also gives (102). □
The same factorization and concatenation arguments apply to full-past rules on C 0 ( R ) \mathcal{C}_0(\mathbb{R}) C 0 ( R ) , using rational coordinates on the whole line. The map x ↦ ( x S + r − x S ) r ∈ R x\mapsto(x_{S+r}-x_S)_{r\in\mathbb{R}} x ↦ ( x S + r − x S ) r ∈ R preserves the two-sided fBm law. They also apply to finite-window rules on [ − M , L ] [-M,L] [ − M , L ] .
Finite-memory approximation
On the two-sided canonical fBm space, compare the finite-window and full-past observation fields, defined for 0 ≤ t ≤ L 0 \le t \le L 0 ≤ t ≤ L by
H t ( M ) : = σ ( B s H : − M ≤ s ≤ t ) ‾ , H t ( − ∞ ) : = σ ( B s H : s ≤ t ) ‾ . \mathcal{H}^{(M)}_t := \overline{\sigma\left(B_s^H:-M \le s \le t\right)}, \qquad\mathcal{H}^{(-\infty)}_t := \overline{\sigma\left(B_s^H:s \le t\right)}. H t ( M ) := σ ( B s H : − M ≤ s ≤ t ) , H t ( − ∞ ) := σ ( B s H : s ≤ t ) .
The full-past field is F t B \mathcal{F}^B_t F t B by lemma A.1. These finite-window fields are right-continuous after completion for t ≥ 0 t \ge0 t ≥ 0 . To see this, set Y r : = B − M + r H − B − M H Y_r := B^H_{-M+r} - B^H_{-M} Y r := B − M + r H − B − M H for 0 ≤ r ≤ M + t 0 \le r \le M+t 0 ≤ r ≤ M + t . Since B 0 H = 0 B^H_0=0 B 0 H = 0 , we have Y M = − B − M H Y_M=-B^H_{-M} Y M = − B − M H , so
B s H = Y s + M − Y M ( − M ≤ s ≤ t ) . B_s^H=Y_{s+M}-Y_M \qquad(-M \le s \le t). B s H = Y s + M − Y M ( − M ≤ s ≤ t ) .
The two observed paths determine each other. The process Y Y Y has the one-sided fBm law, whose completed natural filtration is right-continuous.
Proof of lemma 5.2. Write J L J_L J L for the full-past reset cost. Given η > 0 \eta>0 η > 0 , choose ψ \psi ψ with J L ( ψ ) ≤ u H p a s t ( L ) + η J_L(\psi)\le u_H^{\mathrm{past}}(L)+\eta J L ( ψ ) ≤ u H past ( L ) + η . We approximate this rule by restricting its values to [ − R , R ] [-R,R] [ − R , R ] , discretizing time and then restricting the observation window.
Restricting the control values. First set ψ R : = ( − R ) ∨ ψ ∧ R \psi^R:=(-R)\vee\psi\wedge R ψ R := ( − R ) ∨ ψ ∧ R . Since the map x ↦ ( − R ) ∨ x ∧ R x\mapsto(-R)\vee x\wedge R x ↦ ( − R ) ∨ x ∧ R is 1-Lipschitz and fixes zero,
TV [ 0 , L ] a u g ( ψ R ) ≤ TV [ 0 , L ] a u g ( ψ ) . \operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi^R)\le\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi). TV [ 0 , L ] aug ( ψ R ) ≤ TV [ 0 , L ] aug ( ψ ) .
Finite cost gives E [ ∫ 0 L ψ t 2 d t ] < ∞ \mathbb{E}\left[\int_0^L\psi_t^2\,\mathrm{d}t\right]<\infty E [ ∫ 0 L ψ t 2 d t ] < ∞ and E [ ∣ ψ L ∣ ] < ∞ \mathbb{E}[|\psi_L|]<\infty E [ ∣ ψ L ∣ ] < ∞ . The bounds
∣ ψ t R − B t H ∣ 2 ≤ 2 ∣ ψ t ∣ 2 + 2 ∣ B t H ∣ 2 , |\psi_t^R-B_t^H|^2\le2|\psi_t|^2+2|B_t^H|^2, ∣ ψ t R − B t H ∣ 2 ≤ 2∣ ψ t ∣ 2 + 2∣ B t H ∣ 2 ,
and
∣ ψ L R − B L H ∣ ≤ ∣ ψ L ∣ + ∣ B L H ∣ |\psi_L^R-B_L^H|\le|\psi_L|+|B_L^H| ∣ ψ L R − B L H ∣ ≤ ∣ ψ L ∣ + ∣ B L H ∣
let us apply dominated convergence to the tracking and reset costs. Thus lim sup R J L ( ψ R ) ≤ J L ( ψ ) \limsup_R J_L(\psi^R)\le J_L(\psi) lim sup R J L ( ψ R ) ≤ J L ( ψ ) . Choose R R R with increase at most η \eta η and relabel this bounded rule as ψ \psi ψ .
Time discretization. Next, on a dyadic partition 0 = t 0 < ⋯ < t K = L 0=t_0<\cdots<t_K=L 0 = t 0 < ⋯ < t K = L , define
ψ t π : = ψ t k ( t k ≤ t < t k + 1 ) , ψ L π : = ψ L . \psi_t^\pi:=\psi_{t_k}\quad(t_k\le t<t_{k+1}),\qquad\psi_L^\pi:=\psi_L. ψ t π := ψ t k ( t k ≤ t < t k + 1 ) , ψ L π := ψ L .
Ordered sampling gives
TV [ 0 , L ] a u g ( ψ π ) = ∣ ψ 0 ∣ + ∑ k = 1 K ∣ ψ t k − ψ t k − 1 ∣ ≤ TV [ 0 , L ] a u g ( ψ ) . \operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi^\pi)=|\psi_0|+\sum_{k=1}^{K}|\psi_{t_k}-\psi_{t_{k-1}}|\le\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi). TV [ 0 , L ] aug ( ψ π ) = ∣ ψ 0 ∣ + k = 1 ∑ K ∣ ψ t k − ψ t k − 1 ∣ ≤ TV [ 0 , L ] aug ( ψ ) .
Outside the countable pathwise jump set, ψ t π → ψ t \psi_t^\pi\to\psi_t ψ t π → ψ t as mesh ( π ) → 0 \operatorname{mesh}(\pi)\to0 mesh ( π ) → 0 . Boundedness and dominated convergence give convergence of the tracking cost, while the reset is unchanged. Choose a grid that increases the cost by at most η \eta η , and write ξ k : = ψ t k \xi_k:=\psi_{t_k} ξ k := ψ t k .
Finite observation window. Finally, replace each coordinate by ξ k ( M ) : = E [ ξ k ∣ H t k ( M ) ] \xi_k^{(M)}:=\mathbb{E}[\xi_k\mid\mathcal{H}_{t_k}^{(M)}] ξ k ( M ) := E [ ξ k ∣ H t k ( M ) ] and measure the resulting error by d M : = max 0 ≤ k ≤ K ∥ ξ k ( M ) − ξ k ∥ L 2 d_M:=\max_{0\le k\le K}\|\xi_k^{(M)}-\xi_k\|_{L^2} d M := max 0 ≤ k ≤ K ∥ ξ k ( M ) − ξ k ∥ L 2 . Since, after completion,
⋁ m ∈ N H t k ( m ) = H t k ( − ∞ ) , \bigvee_{m\in\mathbb{N}}\mathcal{H}_{t_k}^{(m)}=\mathcal{H}_{t_k}^{(-\infty)}, m ∈ N ⋁ H t k ( m ) = H t k ( − ∞ ) ,
martingale convergence gives d M → M → ∞ 0 d_M\xrightarrow[M\to\infty]{}0 d M M → ∞ 0 . The coordinates remain bounded by R R R . Their step interpolation ψ π , M \psi^{\pi,M} ψ π , M , including terminal coordinate ξ K ( M ) \xi_K^{(M)} ξ K ( M ) , is a Borel finite-window adapted rule. Since E [ ∫ 0 L ( ψ t π , M − ψ t π ) 2 d t ] ≤ L d M 2 \mathbb{E}\left[\int_0^L(\psi_t^{\pi,M}-\psi_t^\pi)^2\,\mathrm{d}t\right]\le Ld_M^2 E [ ∫ 0 L ( ψ t π , M − ψ t π ) 2 d t ] ≤ L d M 2 , expansion of the square and Cauchy–Schwarz give
∣ E [ ∫ 0 L ( ( ψ t π , M − B t H ) 2 − ( ψ t π − B t H ) 2 ) d t ] ∣ ≤ 2 L d M ( E [ ∫ 0 L ( ψ t π − B t H ) 2 d t ] ) 1 / 2 + L d M 2 → M → ∞ 0. \left|\mathbb{E}\left[\int_0^L\left((\psi_t^{\pi,M}-B_t^H)^2-(\psi_t^\pi-B_t^H)^2\right)\,\mathrm{d}t\right]\right|
\le2\sqrt{L}d_M\left(\mathbb{E}\left[\int_0^L(\psi_t^\pi-B_t^H)^2\,\mathrm{d}t\right]\right)^{1/2}+Ld_M^2\xrightarrow[M\to\infty]{}0. E [ ∫ 0 L ( ( ψ t π , M − B t H ) 2 − ( ψ t π − B t H ) 2 ) d t ] ≤ 2 L d M ( E [ ∫ 0 L ( ψ t π − B t H ) 2 d t ] ) 1/2 + L d M 2 M → ∞ 0.
For each mesh jump,
∣ ∣ ξ k ( M ) − ξ k − 1 ( M ) ∣ − ∣ ξ k − ξ k − 1 ∣ ∣ ≤ ∣ ξ k ( M ) − ξ k ∣ + ∣ ξ k − 1 ( M ) − ξ k − 1 ∣ . \left||\xi_k^{(M)}-\xi_{k-1}^{(M)}|-|\xi_k-\xi_{k-1}|\right|\le|\xi_k^{(M)}-\xi_k|+|\xi_{k-1}^{(M)}-\xi_{k-1}|. ∣ ξ k ( M ) − ξ k − 1 ( M ) ∣ − ∣ ξ k − ξ k − 1 ∣ ≤ ∣ ξ k ( M ) − ξ k ∣ + ∣ ξ k − 1 ( M ) − ξ k − 1 ∣.
Including the initial atom and terminal reset, the expected absolute change in these costs is at most ( 2 K + 2 ) d M → M → ∞ 0 (2K+2)d_M\xrightarrow[M\to\infty]{}0 ( 2 K + 2 ) d M M → ∞ 0 . For large M M M , the resulting rule has cost at most u H p a s t ( L ) + 4 η u_H^{\mathrm{past}}(L)+4\eta u H past ( L ) + 4 η . The finite-memory values decrease in M M M and are bounded below by u H p a s t ( L ) u_H^{\mathrm{past}}(L) u H past ( L ) . Let η → 0 \eta\to0 η → 0 . Removing the terminal penalty gives the same proof for v H p a s t v_H^{\mathrm{past}} v H past .
A continuous-time subadditive lemma
Let f : ( 0 , ∞ ) → [ 0 , ∞ ) f:(0,\infty)\to[0,\infty) f : ( 0 , ∞ ) → [ 0 , ∞ ) be subadditive and locally bounded. Given q > 0 q>0 q > 0 , write L = n q + r L=nq+r L = n q + r , where n : = ⌊ L / q ⌋ ≥ 1 n:=\lfloor L/q\rfloor\ge1 n := ⌊ L / q ⌋ ≥ 1 and 0 ≤ r < q 0\le r<q 0 ≤ r < q . Local boundedness gives C q : = sup q ≤ s ≤ 2 q f ( s ) < ∞ C_q:=\sup_{q\le s\le2q}f(s)<\infty C q := sup q ≤ s ≤ 2 q f ( s ) < ∞ . Instead of using the remainder r r r alone, decompose L = ( n − 1 ) q + ( q + r ) L=(n-1)q+(q+r) L = ( n − 1 ) q + ( q + r ) so that subadditivity only calls on values away from zero:
f ( L ) ≤ ( n − 1 ) f ( q ) + f ( q + r ) ≤ n f ( q ) + C q . f(L)\le(n-1)f(q)+f(q+r)\le nf(q)+C_q. f ( L ) ≤ ( n − 1 ) f ( q ) + f ( q + r ) ≤ n f ( q ) + C q .
As n L → L → ∞ 1 q \frac{n}{L}\xrightarrow[L\to\infty]{}\frac{1}{q} L n L → ∞ q 1 ,
lim sup L → ∞ f ( L ) L ≤ f ( q ) q . \limsup_{L\to\infty}\frac{f(L)}{L}\le\frac{f(q)}{q}. L → ∞ lim sup L f ( L ) ≤ q f ( q ) .
Taking the infimum over q > 0 q>0 q > 0 gives
inf q > 0 f ( q ) q ≤ lim inf L → ∞ f ( L ) L ≤ lim sup L → ∞ f ( L ) L ≤ inf q > 0 f ( q ) q , \inf_{q>0}\frac{f(q)}{q}\le\liminf_{L\to\infty}\frac{f(L)}{L}\le\limsup_{L\to\infty}\frac{f(L)}{L}\le\inf_{q>0}\frac{f(q)}{q}, q > 0 inf q f ( q ) ≤ L → ∞ lim inf L f ( L ) ≤ L → ∞ lim sup L f ( L ) ≤ q > 0 inf q f ( q ) ,
which proves the limit and infimum identity.
The Brownian benchmark
Proof of proposition 4.1. Verification lower bound. Set d : = ( 3 / 2 ) 1 / 3 d:=(3/2)^{1/3} d := ( 3/2 ) 1/3 and κ : = d 2 / 2 \kappa:=d^2/2 κ := d 2 /2 . Define
v : [ t ] R ⟶ R x ⟼ { d 2 x 2 2 − x 4 12 , ∣ x ∣ ≤ d , 5 d 4 12 + ∣ x ∣ − d , ∣ x ∣ > d . v:\begin{aligned}[t]
\mathbb{R}&\longrightarrow\mathbb{R}\\
x&\longmapsto
\begin{cases}
\dfrac{d^2x^2}{2}-\dfrac{x^4}{12},&|x|\le d,\\
\dfrac{5d^4}{12}+|x|-d,&|x|>d.
\end{cases}
\end{aligned} v : [ t ] R x ⟶ R ⟼ ⎩ ⎨ ⎧ 2 d 2 x 2 − 12 x 4 , 12 5 d 4 + ∣ x ∣ − d , ∣ x ∣ ≤ d , ∣ x ∣ > d .
Its derivatives satisfy
v ′ ( x ) = { d 2 x − x 3 / 3 , ∣ x ∣ ≤ d , sign ( x ) , ∣ x ∣ > d , v'(x)=
\begin{cases}
d^2x-x^3/3,&|x|\le d,\\
\operatorname{sign}(x),&|x|>d,
\end{cases} v ′ ( x ) = { d 2 x − x 3 /3 , sign ( x ) , ∣ x ∣ ≤ d , ∣ x ∣ > d ,
and
v ′ ′ ( x ) = { d 2 − x 2 , ∣ x ∣ ≤ d , 0 , ∣ x ∣ > d . v''(x)=
\begin{cases}
d^2-x^2,&|x|\le d,\\
0,&|x|>d.
\end{cases} v ′′ ( x ) = { d 2 − x 2 , 0 , ∣ x ∣ ≤ d , ∣ x ∣ > d .
Because 2 d 3 / 3 = 1 2d^3/3=1 2 d 3 /3 = 1 , both derivatives match at ± d \pm d ± d . Thus v ∈ C 2 ( R ) v\in C^2(\mathbb{R}) v ∈ C 2 ( R ) , ∣ v ′ ∣ ≤ 1 |v'|\le1 ∣ v ′ ∣ ≤ 1 , 0 ≤ v ( x ) ≤ ∣ x ∣ 0\le v(x)\le|x| 0 ≤ v ( x ) ≤ ∣ x ∣ , and
1 2 v ′ ′ ( x ) + 1 2 x 2 − κ = { 0 , ∣ x ∣ ≤ d , ( x 2 − d 2 ) / 2 , ∣ x ∣ > d ≥ 0. \frac{1}{2}v''(x)+\frac{1}{2}x^2-\kappa=
\begin{cases}
0,&|x|\le d,\\
(x^2-d^2)/2,&|x|>d
\end{cases}
\ge0. 2 1 v ′′ ( x ) + 2 1 x 2 − κ = { 0 , ( x 2 − d 2 ) /2 , ∣ x ∣ ≤ d , ∣ x ∣ > d ≥ 0.
For an admissible ψ \psi ψ , put e : = ψ − B e:=\psi-B e := ψ − B and e 0 − : = 0 e_{0-}:=0 e 0 − := 0 . Itô’s formula, including the initial jump, and the Lipschitz bound give
v ( e L ) = v ( e 0 ) − v ( 0 ) + ∫ ( 0 , L ] v ′ ( e t − ) d ψ t c + ∑ 0 < t ≤ L Δ e t ≠ 0 { v ( e t ) − v ( e t − ) } − ∫ 0 L v ′ ( e t − ) d B t + 1 2 ∫ 0 L v ′ ′ ( e t − ) d t ≥ − TV [ 0 , L ] a u g ( ψ ) − ∫ 0 L v ′ ( e t − ) d B t + 1 2 ∫ 0 L v ′ ′ ( e t − ) d t . \begin{aligned}
v(e_L)&=v(e_0)-v(0)+\int_{(0,L]}v'(e_{t-})\,\mathrm{d}\psi_t^c\\
&\quad+\sum_{\substack{0<t\le L\\ \Delta e_t\ne0}}\{v(e_t)-v(e_{t-})\}-\int_0^L v'(e_{t-})\,\mathrm{d}B_t+\frac{1}{2}\int_0^L v''(e_{t-})\,\mathrm{d}t\\
&\ge-\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)-\int_0^L v'(e_{t-})\,\mathrm{d}B_t+\frac{1}{2}\int_0^L v''(e_{t-})\,\mathrm{d}t.
\end{aligned} v ( e L ) = v ( e 0 ) − v ( 0 ) + ∫ ( 0 , L ] v ′ ( e t − ) d ψ t c + 0 < t ≤ L Δ e t = 0 ∑ { v ( e t ) − v ( e t − )} − ∫ 0 L v ′ ( e t − ) d B t + 2 1 ∫ 0 L v ′′ ( e t − ) d t ≥ − TV [ 0 , L ] aug ( ψ ) − ∫ 0 L v ′ ( e t − ) d B t + 2 1 ∫ 0 L v ′′ ( e t − ) d t .
The stochastic integral has mean zero because v ′ v' v ′ is bounded. Taking expectations gives
E [ 1 2 ∫ 0 L e t 2 d t + TV [ 0 , L ] a u g ( ψ ) ] ≥ E [ ∫ 0 L e t 2 + v ′ ′ ( e t ) 2 d t ] − E [ v ( e L ) ] ≥ κ L − E [ v ( e L ) ] . \mathbb{E}\left[\frac{1}{2}\int_0^L e_t^2\,\mathrm{d}t+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)\right]\ge\mathbb{E}\left[\int_0^L\frac{e_t^2+v''(e_t)}{2}\,\mathrm{d}t\right]-\mathbb{E}[v(e_L)]\ge\kappa L-\mathbb{E}[v(e_L)]. E [ 2 1 ∫ 0 L e t 2 d t + TV [ 0 , L ] aug ( ψ ) ] ≥ E [ ∫ 0 L 2 e t 2 + v ′′ ( e t ) d t ] − E [ v ( e L )] ≥ κ L − E [ v ( e L )] .
After adding the reset,
E [ 1 2 ∫ 0 L e t 2 d t + TV [ 0 , L ] a u g ( ψ ) + ∣ e L ∣ ] ≥ κ L + E [ ∣ e L ∣ − v ( e L ) ] ≥ κ L . \mathbb{E}\left[\frac{1}{2}\int_0^L e_t^2\,\mathrm{d}t+\operatorname{TV}^{\mathrm{aug}}_{[0,L]}(\psi)+|e_L|\right]\ge\kappa L+\mathbb{E}[|e_L|-v(e_L)]\ge\kappa L. E [ 2 1 ∫ 0 L e t 2 d t + TV [ 0 , L ] aug ( ψ ) + ∣ e L ∣ ] ≥ κ L + E [ ∣ e L ∣ − v ( e L )] ≥ κ L .
This is the singular-control verification argument [1 , 2 , 3 , 30 ] .
Reflected upper bound. For any d > 0 d > 0 d > 0 , let e = − B + K − − K + e = -B + K^{-} - K^{+} e = − B + K − − K + be reflected on [ − d , d ] [-d,d] [ − d , d ] , starting at zero. The continuous nondecreasing regulators K ± K^{\pm} K ± start at zero, with d K − \mathrm{d}K^{-} d K − supported on { e = − d } \{e=-d\} { e = − d } and d K + \mathrm{d}K^{+} d K + on { e = d } \{e=d\} { e = d } . The control ψ : = K − − K + \psi:= K^{-} - K^{+} ψ := K − − K + has augmented variation T V [ 0 , L ] a u g ( ψ ) = K L − + K L + \mathrm{TV}^{\mathrm{aug}}_{[0,L]}(\psi) = K^{-}_{L} + K^{+}_{L} TV [ 0 , L ] aug ( ψ ) = K L − + K L + . Itô’s formula gives
e L 2 = − 2 ∫ 0 L e t d B t + L − 2 d ( K L − + K L + ) . e_{L}^{2} = -2\int_{0}^{L} e_{t}\,\mathrm{d}B_{t} + L - 2d(K^{-}_{L} + K^{+}_{L}). e L 2 = − 2 ∫ 0 L e t d B t + L − 2 d ( K L − + K L + ) .
Then,
E [ K L − + K L + ] = L − E [ e L 2 ] 2 d = L 2 d + O ( 1 ) . \mathbb{E}[K^{-}_{L} + K^{+}_{L}] = \frac{L-\mathbb{E}[e_{L}^{2}]}{2d} = \frac{L}{2d} + O(1). E [ K L − + K L + ] = 2 d L − E [ e L 2 ] = 2 d L + O ( 1 ) .
To compute the running cost, use
f : ∣ [ − d , d ] ⟶ R x ⟼ x 4 6 − d 2 x 2 3 f:\left|\begin{array}{rcl}
[-d,d] &\longrightarrow& \mathbb{R}\\
x &\longmapsto& \dfrac{x^{4}}{6}-\dfrac{d^{2}x^{2}}{3}
\end{array}\right. f : [ − d , d ] x ⟶ ⟼ R 6 x 4 − 3 d 2 x 2
Its derivatives satisfy
f ′ ( x ) = 2 x ( x 2 − d 2 ) 3 , f ′ ( ± d ) = 0 , 1 2 f ′ ′ ( x ) = x 2 − d 2 3 . f'(x)=\frac{2x(x^{2}-d^{2})}{3},\qquad f'(\pm d)=0,\qquad\frac{1}{2}f''(x)=x^{2}-\frac{d^{2}}{3}. f ′ ( x ) = 3 2 x ( x 2 − d 2 ) , f ′ ( ± d ) = 0 , 2 1 f ′′ ( x ) = x 2 − 3 d 2 .
The regulator terms vanish, so
E [ f ( e L ) ] − f ( 0 ) = E [ ∫ 0 L ( e t 2 − d 2 3 ) d t ] . \mathbb{E}[f(e_{L})]-f(0)=\mathbb{E}\left[\int_{0}^{L}\left(e_{t}^{2}-\frac{d^{2}}{3}\right)\,\mathrm{d}t\right]. E [ f ( e L )] − f ( 0 ) = E [ ∫ 0 L ( e t 2 − 3 d 2 ) d t ] .
Then,
E [ ∫ 0 L e t 2 d t ] = d 2 L 3 + E [ f ( e L ) ] − f ( 0 ) = d 2 L 3 + O ( 1 ) . \mathbb{E}\left[\int_{0}^{L}e_{t}^{2}\,\mathrm{d}t\right]=\frac{d^{2}L}{3}+\mathbb{E}[f(e_{L})]-f(0)=\frac{d^{2}L}{3}+O(1). E [ ∫ 0 L e t 2 d t ] = 3 d 2 L + E [ f ( e L )] − f ( 0 ) = 3 d 2 L + O ( 1 ) .
As ∣ e L ∣ ≤ d |e_{L}|\le d ∣ e L ∣ ≤ d , the reset contributes O ( 1 ) O(1) O ( 1 ) . The cost rate is
c : ∣ ( 0 , ∞ ) ⟶ ( 0 , ∞ ) d ⟼ d 2 6 + 1 2 d c:\left|\begin{array}{rcl}
(0,\infty) &\longrightarrow& (0,\infty)\\
d &\longmapsto& \dfrac{d^{2}}{6}+\dfrac{1}{2d}
\end{array}\right. c : ( 0 , ∞ ) d ⟶ ⟼ ( 0 , ∞ ) 6 d 2 + 2 d 1
Since
c ′ ( d ) = d 3 − 1 2 d 2 , c ′ ′ ( d ) = 1 3 + 1 d 3 > 0 , c'(d)=\frac{d}{3}-\frac{1}{2d^{2}},\qquad c''(d)=\frac{1}{3}+\frac{1}{d^{3}}>0, c ′ ( d ) = 3 d − 2 d 2 1 , c ′′ ( d ) = 3 1 + d 3 1 > 0 ,
the unique minimum occurs at d 3 = 3 / 2 d^{3}=3/2 d 3 = 3/2 , and
c ( d ∗ ) = d ∗ 2 6 + 1 2 d ∗ = d ∗ 2 6 + d ∗ 2 3 = d ∗ 2 2 = κ . c(d_{*})=\frac{d_{*}^{2}}{6}+\frac{1}{2d_{*}}=\frac{d_{*}^{2}}{6}+\frac{d_{*}^{2}}{3}=\frac{d_{*}^{2}}{2}=\kappa. c ( d ∗ ) = 6 d ∗ 2 + 2 d ∗ 1 = 6 d ∗ 2 + 3 d ∗ 2 = 2 d ∗ 2 = κ .
Prediction states
The prediction curve as a state variable
Let W = ( W t ) t ≥ 0 W=(W_{t})_{t\ge0} W = ( W t ) t ≥ 0 be standard Brownian motion with augmented natural filtration F W \mathcal{F}^{W} F W . Choose a jointly Borel, filtration-preserving Volterra kernel K H : { ( t , s ) ∈ [ 0 , ∞ ) 2 : s ≤ t } → R \mathcal{K}_{H}:\{(t,s)\in[0,\infty)^{2}:s\le t\}\to\mathbb{R} K H : {( t , s ) ∈ [ 0 , ∞ ) 2 : s ≤ t } → R [31 ] such that
B t H = ∫ 0 t K H ( t , s ) d W s , F t B H = F t W , (104) B_{t}^{H}=\int_{0}^{t}\mathcal{K}_{H}(t,s)\,\mathrm{d}W_{s},\qquad\mathcal{F}_{t}^{B^{H}}=\mathcal{F}_{t}^{W}, \tag*{(104)} B t H = ∫ 0 t K H ( t , s ) d W s , F t B H = F t W , (104)
with K H ( t , ⋅ ) ∈ L 2 ( [ 0 , t ] ) \mathcal{K}_{H}(t,\cdot)\in L^{2}([0,t]) K H ( t , ⋅ ) ∈ L 2 ([ 0 , t ]) . The state space E : = C l o c ( [ 0 , ∞ ) ; R ) \mathcal{E}:=C_{\mathrm{loc}}([0,\infty);\mathbb{R}) E := C loc ([ 0 , ∞ ) ; R ) is Polish under uniform convergence on compact intervals. Write N E ( g , Q ) \mathcal{N}_{\mathcal{E}}(g,Q) N E ( g , Q ) for the Gaussian probability law on E \mathcal{E} E with mean g ∈ E g\in\mathcal{E} g ∈ E and covariance kernel Q Q Q . For h ≥ 0 h\ge0 h ≥ 0 , the shift
τ h : ∣ E ⟶ E g ⟼ ( r ↦ g ( h + r ) ) \tau_{h}:\left|\begin{array}{rcl}
\mathcal{E} &\longrightarrow& \mathcal{E}\\
g &\longmapsto& (r\mapsto g(h+r))
\end{array}\right. τ h : E g ⟶ ⟼ E ( r ↦ g ( h + r ))
is continuous. The prediction curve is defined by
G t ( r ) : = E [ B t + r H ∣ F t B H ] = ∫ 0 t K H ( t + r , s ) d W s , r ≥ 0. (105) G_t(r) := \mathbb{E}[B^H_{t+r} \mid\mathcal{F}^{B^H}_t] = \int_0^t \mathcal{K}_H(t+r,s)\,\mathrm{d}W_s,\qquad r \ge0. \tag*{(105)} G t ( r ) := E [ B t + r H ∣ F t B H ] = ∫ 0 t K H ( t + r , s ) d W s , r ≥ 0. (105)
The L 2 L^2 L 2 contraction of conditional expectation, followed by Gaussian moment equivalence, gives
E [ ∣ G t ( r ) − G t ( q ) ∣ 2 ] ≤ E [ ∣ B t + r H − B t + q H ∣ 2 ] = ∣ r − q ∣ 2 H . \mathbb{E}\left[|G_t(r)-G_t(q)|^2\right] \le\mathbb{E}\left[|B^H_{t+r}-B^H_{t+q}|^2\right] = |r-q|^{2H}. E [ ∣ G t ( r ) − G t ( q ) ∣ 2 ] ≤ E [ ∣ B t + r H − B t + q H ∣ 2 ] = ∣ r − q ∣ 2 H .
Then,
E [ ∣ G t ( r ) − G t ( q ) ∣ p ] ≤ C p ∣ r − q ∣ p H , p ≥ 2. \mathbb{E}\left[|G_t(r)-G_t(q)|^p\right] \le C_p|r-q|^{pH},\qquad p\ge2. E [ ∣ G t ( r ) − G t ( q ) ∣ p ] ≤ C p ∣ r − q ∣ p H , p ≥ 2.
For p > 1 / H p>1/H p > 1/ H , Kolmogorov’s criterion [32 Chapter 4] gives a continuous version on each compact r r r -interval. Start from the jointly measurable parameterized Wiener integrals for r ∈ Q + r\in\mathbb{Q}_+ r ∈ Q + and extend by continuity. Separability of E \mathcal{E} E gives a version with G t ( 0 ) = B t H G_t(0)=B^H_t G t ( 0 ) = B t H such that
[ 0 , ∞ ) × Ω ⟶ E ( t , ω ) ⟼ G t ( ω ) \begin{aligned}
[0,\infty)\times\Omega&\longrightarrow\mathcal{E}\\
(t,\omega)&\longmapsto G_t(\omega)
\end{aligned} [ 0 , ∞ ) × Ω ( t , ω ) ⟶ E ⟼ G t ( ω )
is B ( [ 0 , ∞ ) ) ⊗ F \mathcal{B}([0,\infty))\otimes\mathcal{F} B ([ 0 , ∞ )) ⊗ F –B ( E ) \mathcal{B}(\mathcal{E}) B ( E ) measurable.
Theorem G.1 (Markov property of the prediction state). The [ 0 , ∞ ) × E [0,\infty)\times\mathcal{E} [ 0 , ∞ ) × E -valued process ( ( t , G t ) ) t ≥ 0 ((t,G_t))_{t\ge0} (( t , G t ) ) t ≥ 0 is Markov. For h , r ≥ 0 h,r\ge0 h , r ≥ 0 ,
G t + h ( r ) = G t ( h + r ) + ∫ t t + h K H ( t + h + r , s ) d W s , (106) G_{t+h}(r)=G_t(h+r)+\int_t^{t+h}\mathcal{K}_H(t+h+r,s)\,\mathrm{d}W_s, \tag*{(106)} G t + h ( r ) = G t ( h + r ) + ∫ t t + h K H ( t + h + r , s ) d W s , (106) and, conditionally on F t B H \mathcal{F}^{B^H}_t F t B H ,
B t + r H = G t ( r ) + ∫ t t + r K H ( t + r , s ) d W s . (107) B^H_{t+r}=G_t(r)+\int_t^{t+r}\mathcal{K}_H(t+r,s)\,\mathrm{d}W_s. \tag*{(107)} B t + r H = G t ( r ) + ∫ t t + r K H ( t + r , s ) d W s . (107) With Q t Q_t Q t defined in (108) ,
L ( ( B t + r H ) r ≥ 0 ∣ F t B H ) = N E ( G t , Q t ) . \mathcal{L}\left((B^H_{t+r})_{r\ge0}\mid\mathcal{F}^{B^H}_t\right)=\mathcal{N}_{\mathcal{E}}(G_t,Q_t). L ( ( B t + r H ) r ≥ 0 ∣ F t B H ) = N E ( G t , Q t ) .
Proof. Splitting the Wiener integral at t t t gives
B t + r H = ∫ 0 t K H ( t + r , s ) d W s + ∫ t t + r K H ( t + r , s ) d W s = G t ( r ) + ∫ t t + r K H ( t + r , s ) d W s . \begin{aligned}
B^H_{t+r}
&=\int_0^t\mathcal{K}_H(t+r,s)\,\mathrm{d}W_s+\int_t^{t+r}\mathcal{K}_H(t+r,s)\,\mathrm{d}W_s\\
&=G_t(r)+\int_t^{t+r}\mathcal{K}_H(t+r,s)\,\mathrm{d}W_s.
\end{aligned} B t + r H = ∫ 0 t K H ( t + r , s ) d W s + ∫ t t + r K H ( t + r , s ) d W s = G t ( r ) + ∫ t t + r K H ( t + r , s ) d W s .
The second term is centered Gaussian and independent of F t B H = F t W \mathcal{F}^{B^H}_t=\mathcal{F}^W_t F t B H = F t W . Its covariance is
Q t ( r , q ) : = ∫ t t + ( r ∧ q ) K H ( t + r , s ) K H ( t + q , s ) d s . (108) Q_t(r,q):=\int_t^{t+(r\wedge q)}\mathcal{K}_H(t+r,s)\mathcal{K}_H(t+q,s)\,\mathrm{d}s. \tag*{(108)} Q t ( r , q ) := ∫ t t + ( r ∧ q ) K H ( t + r , s ) K H ( t + q , s ) d s . (108)
Hence
L ( ( B t + r H ) r ≥ 0 ∣ F t B H ) = N E ( G t , Q t ) . \mathcal{L}\left((B^H_{t+r})_{r\ge0}\mid\mathcal{F}^{B^H}_t\right)=\mathcal{N}_{\mathcal{E}}(G_t,Q_t). L ( ( B t + r H ) r ≥ 0 ∣ F t B H ) = N E ( G t , Q t ) .
Likewise,
G t + h ( r ) = ∫ 0 t K H ( t + h + r , s ) d W s + ∫ t t + h K H ( t + h + r , s ) d W s = G t ( h + r ) + ∫ t t + h K H ( t + h + r , s ) d W s . \begin{aligned}
G_{t+h}(r)
&=\int_0^t\mathcal{K}_H(t+h+r,s)\,\mathrm{d}W_s+\int_t^{t+h}\mathcal{K}_H(t+h+r,s)\,\mathrm{d}W_s\\
&=G_t(h+r)+\int_t^{t+h}\mathcal{K}_H(t+h+r,s)\,\mathrm{d}W_s.
\end{aligned} G t + h ( r ) = ∫ 0 t K H ( t + h + r , s ) d W s + ∫ t t + h K H ( t + h + r , s ) d W s = G t ( h + r ) + ∫ t t + h K H ( t + h + r , s ) d W s .
Denote the last integral, as a function of r r r , by I t , h I_{t,h} I t , h . The Gaussian increment estimate gives a continuous version I t , h ∈ L 0 ( Ω ; E ) I_{t,h}\in L^0(\Omega;\mathcal{E}) I t , h ∈ L 0 ( Ω ; E ) . It is independent of F t B H \mathcal{F}^{B^H}_t F t B H , centered Gaussian, and has covariance
E [ I t , h ( r ) I t , h ( q ) ] = ∫ t t + h K H ( t + h + r , s ) K H ( t + h + q , s ) d s . \mathbb{E}[I_{t,h}(r)I_{t,h}(q)]=\int_t^{t+h}\mathcal{K}_H(t+h+r,s)\mathcal{K}_H(t+h+q,s)\,\mathrm{d}s. E [ I t , h ( r ) I t , h ( q )] = ∫ t t + h K H ( t + h + r , s ) K H ( t + h + q , s ) d s .
Define the transition map
P t , h : E ⟶ P ( E ) g ⟼ L ( τ h g + I t , h ) \begin{aligned}
P_{t,h}:\quad\mathcal{E}&\longrightarrow\mathcal{P}(\mathcal{E})\\
g&\longmapsto\mathcal{L}(\tau_hg+I_{t,h})
\end{aligned} P t , h : E g ⟶ P ( E ) ⟼ L ( τ h g + I t , h )
For every A ∈ B ( E ) A \in\mathcal{B}(E) A ∈ B ( E ) ,
P t , h ( g , A ) = E [ 1 A ( τ h g + I t , h ) ] . P_{t,h}(g,A)=\mathbb{E}[1_{A}(\tau_{h}g+I_{t,h})]. P t , h ( g , A ) = E [ 1 A ( τ h g + I t , h )] .
Since addition and τ h \tau_{h} τ h are continuous, P t , h P_{t,h} P t , h is a Borel kernel. Independence gives
L ( G t + h ∣ F t B H ) = P t , h ( G t , ⋅ ) . \mathcal{L}(G_{t+h}\mid\mathcal{F}^{B^{H}}_{t})=P_{t,h}(G_{t},\cdot). L ( G t + h ∣ F t B H ) = P t , h ( G t , ⋅ ) .
Together with the deterministic update t ↦ t + h t\mapsto t+h t ↦ t + h , this proves the Markov property of ( t , G t ) (t,G_{t}) ( t , G t ) .
Corollary G.2 (prediction state for continuation values). Fix t ≥ 0 t\geq0 t ≥ 0 , τ > 0 \tau>0 τ > 0 , and λ > 0 \lambda>0 λ > 0 , and write E t [ ⋅ ] : = E [ ⋅ ∣ F t B H ] \mathbb{E}_{t}[\cdot]:=\mathbb{E}[\cdot\mid\mathcal{F}^{B^{H}}_{t}] E t [ ⋅ ] := E [ ⋅ ∣ F t B H ] . Let a ˉ , b ˉ : [ 0 , τ ] → ( 0 , ∞ ) \bar{a},\bar{b}:[0,\tau]\to(0,\infty) a ˉ , b ˉ : [ 0 , τ ] → ( 0 , ∞ ) be bounded Borel functions bounded away from zero. For x ∈ R x\in\mathbb{R} x ∈ R , let A t , t + τ ( x ) \mathcal{A}_{t,t+\tau}(x) A t , t + τ ( x ) be the set of F B H \mathcal{F}^{B^{H}} F B H -adapted right-continuous finite-variation controls ξ \xi ξ on [ t , t + τ ] [t,t+\tau] [ t , t + τ ] with pre-initial value ξ t − = x \xi_{t-}=x ξ t − = x and finite expected cost. The variation on this interval includes the atom ∣ ξ t − x ∣ |\xi_{t}-x| ∣ ξ t − x ∣ . Define
V t , τ ( x ) : = ess inf ξ ∈ A t , t + τ ( x ) E t [ 1 2 ∫ t t + τ a ˉ s − t ( ξ s − B s H ) 2 d s + λ ∫ [ t , t + τ ] b ˉ s − t d ∣ ξ ∣ s ] . (109) \mathcal{V}_{t,\tau}(x):=\operatorname*{ess\,inf}_{\xi\in\mathcal{A}_{t,t+\tau}(x)}
\mathbb{E}_{t}\left[\frac{1}{2}\int_{t}^{t+\tau}\bar{a}_{s-t}(\xi_{s}-B^{H}_{s})^{2}\,\mathrm{d}s+\lambda\int_{[t,t+\tau]}\bar{b}_{s-t}\,\mathrm{d}|\xi|_{s}\right]. \tag*{(109)} V t , τ ( x ) := ξ ∈ A t , t + τ ( x ) ess inf E t [ 2 1 ∫ t t + τ a ˉ s − t ( ξ s − B s H ) 2 d s + λ ∫ [ t , t + τ ] b ˉ s − t d ∣ ξ ∣ s ] . (109)
There is a Borel function V t , λ , τ a ˉ , b ˉ \mathcal{V}^{\bar{a},\bar{b}}_{t,\lambda,\tau} V t , λ , τ a ˉ , b ˉ on E × R \mathcal{E}\times\mathbb{R} E × R , with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , such that, for every x ∈ R x\in\mathbb{R} x ∈ R ,
V t , τ ( x ) = V t , λ , τ a ˉ , b ˉ ( G t , x ) P-a.s. (110) \mathcal{V}_{t,\tau}(x)=\mathcal{V}^{\bar{a},\bar{b}}_{t,\lambda,\tau}(G_{t},x)\quad\text{P-a.s.} \tag*{(110)} V t , τ ( x ) = V t , λ , τ a ˉ , b ˉ ( G t , x ) P-a.s. (110)
Since G t ( 0 ) = B t H G_{t}(0)=B^{H}_{t} G t ( 0 ) = B t H , the same value can be written in the variables ( G t , e ) (G_{t},e) ( G t , e ) with e : = x − G t ( 0 ) e:=x-G_{t}(0) e := x − G t ( 0 ) and x = e + G t ( 0 ) x=e+G_{t}(0) x = e + G t ( 0 ) .
Proof. Introduce the future innovation I t : = ( B t + s H − G t ( s ) ) 0 ≤ s ≤ τ I^{t}:=(B^{H}_{t+s}-G_{t}(s))_{0\leq s\leq\tau} I t := ( B t + s H − G t ( s ) ) 0 ≤ s ≤ τ and the observed past P t : = ( B r H ) 0 ≤ r ≤ t P_{t}:=(B^{H}_{r})_{0\leq r\leq t} P t := ( B r H ) 0 ≤ r ≤ t . These are path-valued random variables in C ( [ 0 , τ ] ; R ) \mathcal{C}([0,\tau];\mathbb{R}) C ([ 0 , τ ] ; R ) and C ( [ 0 , t ] ; R ) \mathcal{C}([0,t];\mathbb{R}) C ([ 0 , t ] ; R ) , respectively. By theorem G.1, I t I^{t} I t is independent of F t B H \mathcal{F}^{B^{H}}_{t} F t B H and has covariance Q t Q_{t} Q t . Let H t \mathbb{H}^{t} H t be the usual augmentation of H s t , 0 : = σ ( G t , I r t : 0 ≤ r ≤ s ) \mathcal{H}^{t,0}_{s}:=\sigma(G_{t},I^{t}_{r}:0\leq r\leq s) H s t , 0 := σ ( G t , I r t : 0 ≤ r ≤ s ) . Brownian independence and the zero–one law imply
H 0 t , 0 = σ ( G t ) P , F t + s B H = H s t ∨ σ ( P t ) P , P t ⊥ ⊥ H t ∣ G t . \mathcal{H}^{t,0}_{0}=\sigma(G_{t})^{\mathbb{P}},\qquad
\mathcal{F}^{B^{H}}_{t+s}=\mathcal{H}^{t}_{s}\vee\sigma(P_{t})^{\mathbb{P}},\qquad
P_{t}\perp\!\!\!\perp\mathcal{H}^{t}\mid G_{t}. H 0 t , 0 = σ ( G t ) P , F t + s B H = H s t ∨ σ ( P t ) P , P t ⊥ ⊥ H t ∣ G t .
To apply (48) , center the target and initial position at G t ( 0 ) G_{t}(0) G t ( 0 ) by setting X s : = G t ( s ) − G t ( 0 ) + I s t X_{s}:=G_{t}(s)-G_{t}(0)+I^{t}_{s} X s := G t ( s ) − G t ( 0 ) + I s t and x 0 : = x − G t ( 0 ) x_{0}:=x-G_{t}(0) x 0 := x − G t ( 0 ) . Then X 0 = 0 X_{0}=0 X 0 = 0 and x 0 ∈ L 0 ( Ω , H 0 t , P ) x_{0}\in L^{0}(\Omega,\mathcal{H}^{t}_{0},\mathbb{P}) x 0 ∈ L 0 ( Ω , H 0 t , P ) . The weighted continuation cost is unchanged. The conditional initial-information identity, applied with the auxiliary variable P t P_{t} P t , shows that V t , τ ( x ) \mathcal{V}_{t,\tau}(x) V t , τ ( x ) is measurable with respect to σ ( G t ) \sigma(G_{t}) σ ( G t ) . The Doob–Dynkin lemma [32 ] gives a Borel function V x V_{x} V x on E \mathcal{E} E , with values in [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , satisfying V t , τ ( x ) = V x ( G t ) \mathcal{V}_{t,\tau}(x)=V_{x}(G_{t}) V t , τ ( x ) = V x ( G t ) almost surely. For any x , y ∈ R x,y\in\mathbb{R} x , y ∈ R , the same control is admissible for both positions, and its initial atoms differ by at most ∣ x − y ∣ |x-y| ∣ x − y ∣ . Thus
∣ V t , τ ( x ) − V t , τ ( y ) ∣ ≤ λ b ˉ 0 ∣ x − y ∣ a.s. |\mathcal{V}_{t,\tau}(x)-\mathcal{V}_{t,\tau}(y)|\leq\lambda\bar{b}_{0}|x-y|\quad\text{a.s.} ∣ V t , τ ( x ) − V t , τ ( y ) ∣ ≤ λ b ˉ 0 ∣ x − y ∣ a.s.
Let E 0 ⊂ E E_{0}\subset\mathcal{E} E 0 ⊂ E be the Borel full-L ( G t ) \mathcal{L}(G_{t}) L ( G t ) -measure set on which all rational values V q V_{q} V q are finite and obey this bound. Define
V t , λ , τ a ˉ , b ˉ : { E × R ⟶ [ 0 , ∞ ) ( g , x ) ⟼ { inf q ∈ Q { V q ( g ) + λ b ˉ 0 ∣ x − q ∣ } , g ∈ E 0 , 0 , g ∉ E 0 . \mathcal{V}^{\bar{a},\bar{b}}_{t,\lambda,\tau}:
\left\{
\begin{aligned}
\mathcal{E}\times\mathbb{R}&\longrightarrow[0,\infty)\\
(g,x)&\longmapsto
\begin{cases}
\inf_{q\in\mathbb{Q}}\{V_{q}(g)+\lambda\bar{b}_{0}|x-q|\},&g\in E_{0},\\
0,&g\notin E_{0}.
\end{cases}
\end{aligned}
\right. V t , λ , τ a ˉ , b ˉ : ⎩ ⎨ ⎧ E × R ( g , x ) ⟶ [ 0 , ∞ ) ⟼ { inf q ∈ Q { V q ( g ) + λ b ˉ 0 ∣ x − q ∣ } , 0 , g ∈ E 0 , g ∈ / E 0 .
This countable infimum is Borel and finite. For g ∈ E 0 g\in E_{0} g ∈ E 0 , it is the unique λ b ˉ 0 \lambda\bar{b}_{0} λ b ˉ 0 -Lipschitz extension of the rational values. Approximating any fixed x x x by rationals proves (109).
Remark G.3. The state is generally infinite-dimensional. See also [19 , 22 , 23 , 24 ] .
Why the scalar error is not a general control state
The next one-step problem shows that a quadratic tracking decision can use information beyond the current error. For fractional Gaussian noise Δ k : = B k H − B k − 1 H \Delta_{k}:=B^{H}_{k}-B^{H}_{k-1} Δ k := B k H − B k − 1 H , the increment covariances are
γ k : = Cov ( Δ 1 , Δ k + 1 ) = 1 2 ( ( k + 1 ) 2 H − 2 k 2 H + ∣ k − 1 ∣ 2 H ) . \gamma_{k}:=\operatorname{Cov}(\Delta_{1},\Delta_{k+1})=\frac{1}{2}\left((k+1)^{2H}-2k^{2H}+|k-1|^{2H}\right). γ k := Cov ( Δ 1 , Δ k + 1 ) = 2 1 ( ( k + 1 ) 2 H − 2 k 2 H + ∣ k − 1 ∣ 2 H ) .
Proposition G.4 (a one-step test of scalar-state sufficiency). Fix a , b , λ > 0 a,b,\lambda>0 a , b , λ > 0 and u ∈ R u\in\mathbb{R} u ∈ R . For each observation ( Δ 1 , Δ 2 ) (\Delta_{1},\Delta_{2}) ( Δ 1 , Δ 2 ) , let v ∗ v^{*} v ∗ be the unique minimizer over v ∈ R v\in\mathbb{R} v ∈ R of the conditional objective
a 2 E [ ( v − B 3 H ) 2 ∣ Δ 1 , Δ 2 ] + λ b ∣ v − u ∣ . (111) \frac{a}{2}\mathbb{E}\left[(v-B^{H}_{3})^{2}\mid\Delta_{1},\Delta_{2}\right]+\lambda b|v-u|. \tag*{(111)} 2 a E [ ( v − B 3 H ) 2 ∣ Δ 1 , Δ 2 ] + λb ∣ v − u ∣. (111)
The minimizer satisfies v ∗ ∈ L 0 ( Ω , σ ( Δ 1 , Δ 2 ) , P ) v^{*}\in L^{0}(\Omega,\sigma(\Delta_{1},\Delta_{2}),\mathbb{P}) v ∗ ∈ L 0 ( Ω , σ ( Δ 1 , Δ 2 ) , P ) . If H ≠ 1 / 2 H\neq1/2 H = 1/2 , v ∗ v^{*} v ∗ is not σ ( B 2 H ) \sigma(B^{H}_{2}) σ ( B 2 H ) -measurable and therefore cannot be a function of u − B 2 H u-B^{H}_{2} u − B 2 H alone. Proof. Let B ^ 3 : = E [ B 3 H ∣ Δ 1 , Δ 2 ] \widehat{B}_3 := \mathbb{E}[B_3^H \mid\Delta_1,\Delta_2] B 3 := E [ B 3 H ∣ Δ 1 , Δ 2 ] . The identity
E [ ( v − B 3 H ) 2 ∣ Δ 1 , Δ 2 ] = ( v − B ^ 3 ) 2 + Var ( B 3 H ∣ Δ 1 , Δ 2 ) \mathbb{E}[(v-B_3^H)^2 \mid\Delta_1,\Delta_2] = (v-\widehat{B}_3)^2+\operatorname{Var}(B_3^H \mid\Delta_1,\Delta_2) E [( v − B 3 H ) 2 ∣ Δ 1 , Δ 2 ] = ( v − B 3 ) 2 + Var ( B 3 H ∣ Δ 1 , Δ 2 )
reduces the problem to the strictly convex function a ( v − B ^ 3 ) 2 / 2 + λ b ∣ v − u ∣ a(v-\widehat{B}_3)^2/2+\lambda b|v-u| a ( v − B 3 ) 2 /2 + λb ∣ v − u ∣ . Its subgradient condition is
0 ∈ a ( v − B ^ 3 ) + λ b ∂ ∣ v − u ∣ , ∂ ∣ v − u ∣ = { { − 1 } , v < u , [ − 1 , 1 ] , v = u , { 1 } , v > u . 0 \in a(v-\widehat{B}_3)+\lambda b\partial|v-u|,\qquad
\partial|v-u| =
\begin{cases}
\{-1\}, & v<u,\\
[-1,1], & v=u,\\
\{1\}, & v>u.
\end{cases} 0 ∈ a ( v − B 3 ) + λb ∂ ∣ v − u ∣ , ∂ ∣ v − u ∣ = ⎩ ⎨ ⎧ { − 1 } , [ − 1 , 1 ] , { 1 } , v < u , v = u , v > u .
Solving on v < u v<u v < u , v = u v=u v = u , and v > u v>u v > u gives
v ∗ = u + sign ( B ^ 3 − u ) ( ∣ B ^ 3 − u ∣ − λ b a ) + . (112) v^* = u+\operatorname{sign}(\widehat{B}_3-u)\left(|\widehat{B}_3-u|-\frac{\lambda b}{a}\right)_+ . \tag*{(112)} v ∗ = u + sign ( B 3 − u ) ( ∣ B 3 − u ∣ − a λb ) + . (112)
Put r : = γ 1 = 2 2 H − 1 − 1 ∈ ( − 1 / 2 , 1 ) r:=\gamma_1=2^{2H-1}-1\in(-1/2,1) r := γ 1 = 2 2 H − 1 − 1 ∈ ( − 1/2 , 1 ) and s : = γ 2 s:=\gamma_2 s := γ 2 . Stationarity identifies the observed covariance matrix and its cross-covariance with the next increment:
Cov ( ( Δ 1 , Δ 2 ) ) = ( 1 r r 1 ) , Cov ( Δ 3 , ( Δ 1 , Δ 2 ) ) = ( s , r ) . \operatorname{Cov}((\Delta_1,\Delta_2))=
\begin{pmatrix}
1&r\\
r&1
\end{pmatrix},
\qquad
\operatorname{Cov}(\Delta_3,(\Delta_1,\Delta_2))=(s,r). Cov (( Δ 1 , Δ 2 )) = ( 1 r r 1 ) , Cov ( Δ 3 , ( Δ 1 , Δ 2 )) = ( s , r ) .
Gaussian regression therefore yields
E [ Δ 3 ∣ Δ 1 , Δ 2 ] = 1 1 − r 2 ( s , r ) ( 1 − r − r 1 ) ( Δ 1 Δ 2 ) = s − r 2 1 − r 2 Δ 1 + r ( 1 − s ) 1 − r 2 Δ 2 . \begin{aligned}
\mathbb{E}[\Delta_3\mid\Delta_1,\Delta_2]
=\frac{1}{1-r^2}(s,r)
\begin{pmatrix}
1&-r\\
-r&1
\end{pmatrix}
\begin{pmatrix}
\Delta_1\\
\Delta_2
\end{pmatrix} \\
&=\frac{s-r^2}{1-r^2}\Delta_1+\frac{r(1-s)}{1-r^2}\Delta_2.
\end{aligned} E [ Δ 3 ∣ Δ 1 , Δ 2 ] = 1 − r 2 1 ( s , r ) ( 1 − r − r 1 ) ( Δ 1 Δ 2 ) = 1 − r 2 s − r 2 Δ 1 + 1 − r 2 r ( 1 − s ) Δ 2 .
Using B 2 H = Δ 1 + Δ 2 B_2^H=\Delta_1+\Delta_2 B 2 H = Δ 1 + Δ 2 and
s − r 2 − r ( 1 − s ) 1 − r 2 = ( s − r ) ( 1 + r ) ( 1 − r ) ( 1 + r ) = s − r 1 − r , \frac{s-r^2-r(1-s)}{1-r^2}
=\frac{(s-r)(1+r)}{(1-r)(1+r)}
=\frac{s-r}{1-r}, 1 − r 2 s − r 2 − r ( 1 − s ) = ( 1 − r ) ( 1 + r ) ( s − r ) ( 1 + r ) = 1 − r s − r ,
we obtain
B ^ 3 = ( 1 + r ( 1 − s ) 1 − r 2 ) B 2 H + s − r 1 − r Δ 1 . (113) \widehat{B}_3=\left(1+\frac{r(1-s)}{1-r^2}\right)B_2^H+\frac{s-r}{1-r}\Delta_1. \tag*{(113)} B 3 = ( 1 + 1 − r 2 r ( 1 − s ) ) B 2 H + 1 − r s − r Δ 1 . (113)
For α : = 2 H ∈ ( 0 , 2 ) \alpha:=2H\in(0,2) α := 2 H ∈ ( 0 , 2 ) , define
d : [ 0 , ∞ ) ⟶ R x ⟼ ( x + 1 ) α − x α d:\quad
\begin{array}{rcl}
[0,\infty)&\longrightarrow&\mathbb{R}\\
x&\longmapsto&(x+1)^\alpha-x^\alpha
\end{array} d : [ 0 , ∞ ) x ⟶ ⟼ R ( x + 1 ) α − x α
For k ≥ 1 k\geq1 k ≥ 1 and x > 0 x>0 x > 0 , its increments and curvature satisfy
2 γ k = d ( k ) − d ( k − 1 ) , d ′ ′ ( x ) = α ( α − 1 ) ( ( x + 1 ) α − 2 − x α − 2 ) . 2\gamma_k=d(k)-d(k-1),\qquad
d''(x)=\alpha(\alpha-1)\left((x+1)^{\alpha-2}-x^{\alpha-2}\right). 2 γ k = d ( k ) − d ( k − 1 ) , d ′′ ( x ) = α ( α − 1 ) ( ( x + 1 ) α − 2 − x α − 2 ) .
The bracket is negative. Hence d d d is strictly convex for α < 1 \alpha<1 α < 1 and strictly concave for α > 1 \alpha>1 α > 1 , so
γ 2 > γ 1 ( H < 1 / 2 ) , γ 2 < γ 1 ( H > 1 / 2 ) . \gamma_2>\gamma_1\quad(H<1/2),\qquad
\gamma_2<\gamma_1\quad(H>1/2). γ 2 > γ 1 ( H < 1/2 ) , γ 2 < γ 1 ( H > 1/2 ) .
Thus s − r ≠ 0 s-r\ne0 s − r = 0 for H ≠ 1 / 2 H\ne1/2 H = 1/2 . Finally,
Var ( B 2 H ) = 2 ( 1 + r ) , Cov ( Δ 1 , B 2 H ) = 1 + r . \operatorname{Var}(B_2^H)=2(1+r),\qquad
\operatorname{Cov}(\Delta_1,B_2^H)=1+r. Var ( B 2 H ) = 2 ( 1 + r ) , Cov ( Δ 1 , B 2 H ) = 1 + r .
Then,
Var ( Δ 1 ∣ B 2 H ) = 1 − ( 1 + r ) 2 2 ( 1 + r ) = 1 − r 2 . \operatorname{Var}(\Delta_1\mid B_2^H)
=1-\frac{(1+r)^2}{2(1+r)}
=\frac{1-r}{2}. Var ( Δ 1 ∣ B 2 H ) = 1 − 2 ( 1 + r ) ( 1 + r ) 2 = 2 1 − r .
Then,
Var ( B ^ 3 ∣ B 2 H ) = ( s − r 1 − r ) 2 1 − r 2 = ( s − r ) 2 2 ( 1 − r ) > 0. (114) \operatorname{Var}(\widehat{B}_3\mid B_2^H)
=\left(\frac{s-r}{1-r}\right)^2\frac{1-r}{2}
=\frac{(s-r)^2}{2(1-r)}>0. \tag*{(114)} Var ( B 3 ∣ B 2 H ) = ( 1 − r s − r ) 2 2 1 − r = 2 ( 1 − r ) ( s − r ) 2 > 0. (114)
The conditional Gaussian law of B ^ 3 \widehat{B}_3 B 3 therefore has full support. The soft-threshold map in (112) has slope one on both tails, each of positive conditional probability. Thus Var ( v ∗ ∣ B 2 H ) > 0 \operatorname{Var}(v^*\mid B_2^H)>0 Var ( v ∗ ∣ B 2 H ) > 0 , excluding σ ( B 2 H ) \sigma(B_2^H) σ ( B 2 H ) -measurability and, for fixed u u u , any representation through u − B 2 H u-B_2^H u − B 2 H alone.
Technical estimates for local homogenization
Stability of reset values
Proof of lemma 6.3. The same control ψ ∈ A L ( G ) \psi\in\mathcal{A}_L(\mathrm{G}) ψ ∈ A L ( G ) is admissible for both targets. The variation terms cancel, so we can expand the squares and apply the triangle inequality. Cauchy–Schwarz then controls the mixed term.
J L X ( ψ ) − J L Y ( ψ ) = E [ ∫ 0 L { − ( ψ t − Y t ) ( X t − Y t ) + 1 2 ∣ X t − Y t ∣ 2 } d t ] + E [ ∣ ψ L − X L ∣ − ∣ ψ L − Y L ∣ ] ≤ E [ ∫ 0 L ∣ ψ t − Y t ∣ ∣ X t − Y t ∣ d t ] + A L 2 + B L ≤ ( E [ ∫ 0 L ( ψ t − Y t ) 2 d t ] ) 1 / 2 A L + A L 2 + B L ≤ 2 J L Y ( ψ ) A L + A L 2 + B L . \begin{aligned}
J_L^X(\psi)-J_L^Y(\psi)=\mathbb{E}\left[\int_0^L\left\{-(\psi_t-Y_t)(X_t-Y_t)+\frac{1}{2}|X_t-Y_t|^2\right\}\,\mathrm{d}t\right]
+\mathbb{E}\left[|\psi_L-X_L|-|\psi_L-Y_L|\right] \\
&\leq\mathbb{E}\left[\int_0^L|\psi_t-Y_t||X_t-Y_t|\,\mathrm{d}t\right]+\frac{A_L}{2}+B_L \\
&\leq\left(\mathbb{E}\left[\int_0^L(\psi_t-Y_t)^2\,\mathrm{d}t\right]\right)^{1/2}\sqrt{A_L}+\frac{A_L}{2}+B_L \\
&\leq\sqrt{2J_L^Y(\psi)A_L}+\frac{A_L}{2}+B_L.
\end{aligned} J L X ( ψ ) − J L Y ( ψ ) = E [ ∫ 0 L { − ( ψ t − Y t ) ( X t − Y t ) + 2 1 ∣ X t − Y t ∣ 2 } d t ] + E [ ∣ ψ L − X L ∣ − ∣ ψ L − Y L ∣ ] ≤ E [ ∫ 0 L ∣ ψ t − Y t ∣∣ X t − Y t ∣ d t ] + 2 A L + B L ≤ ( E [ ∫ 0 L ( ψ t − Y t ) 2 d t ] ) 1/2 A L + 2 A L + B L ≤ 2 J L Y ( ψ ) A L + 2 A L + B L .
Taking a minimizing sequence for Y Y Y gives
U X − U Y ≤ 2 U Y A L + A L 2 + B L . U_X-U_Y\leq\sqrt{2U_YA_L}+\frac{A_L}{2}+B_L. U X − U Y ≤ 2 U Y A L + 2 A L + B L .
Interchange X , Y X,Y X , Y and apply the bound from the smaller value to the larger to obtain (42). Moreover, Young’s inequality gives
2 U Y A L ≤ δ U Y + A L 2 δ \sqrt{2U_YA_L}\leq\delta U_Y+\frac{A_L}{2\delta} 2 U Y A L ≤ δ U Y + 2 δ A L
and hence (42) . Finally, if U X , U Y ≤ C L U_X,U_Y\leq CL U X , U Y ≤ C L , divide by L L L to get
∣ U X − U Y ∣ L ≤ B L L + A L 2 L + 2 C A L L → L → ∞ 0. \frac{|U_X-U_Y|}{L}\leq\frac{B_L}{L}+\frac{A_L}{2L}+\sqrt{\frac{2CA_L}{L}}\xrightarrow[L\to\infty]{}0. L ∣ U X − U Y ∣ ≤ L B L + 2 L A L + L 2 C A L L → ∞ 0.
Proof of lemma 6.5. We first choose a horizon on which the tangent error remains uniform. Put d J : = T − sup J > 0 d_J:=T-\sup J>0 d J := T − sup J > 0 . For integers m ≥ 2 m\geq2 m ≥ 2 and 0 < h < min { h ∗ , d J / ( c + + m 2 ) } 0<h<\min\{h_\ast,d_J/(c_++m^2)\} 0 < h < min { h ∗ , d J / ( c + + m 2 )} , define
ε m : ( 0 , min { h ∗ , d J c + + m 2 } ) ⟶ [ 0 , ∞ ) \varepsilon_m:\left(0,\min\left\{h_\ast,\frac{d_J}{c_++m^2}\right\}\right)\longrightarrow[0,\infty) ε m : ( 0 , min { h ∗ , c + + m 2 d J } ) ⟶ [ 0 , ∞ )
h ⟼ sup t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ m 2 ∥ X u t , h , c − B u t , h , c ∥ L 2 ( P ) . h\longmapsto\sup_{\substack{t\in J,\ c\in[c_-,c_+]\\0\leq u\leq m^2}}\left\|X_u^{t,h,c}-B_u^{t,h,c}\right\|_{L^2(\mathbb{P})}. h ⟼ t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ m 2 sup X u t , h , c − B u t , h , c L 2 ( P ) .
By (FT2), ε m ( h ) → h → 0 0 \varepsilon_m(h)\xrightarrow[h\to0]{}0 ε m ( h ) h → 0 0 for each fixed m m m . Starting with any h 1 > 0 h_1>0 h 1 > 0 , choose recursively
0 < h m < min { h m − 1 2 , h ∗ 2 , 1 m 2 , d J 2 c + + m 2 } , sup 0 < h ≤ h m ε m ( h ) ≤ m − 2 . 0<h_m<\min\left\{\frac{h_{m-1}}{2},\frac{h_\ast}{2},\frac{1}{m^2},\frac{d_J}{2c_++m^2}\right\},\qquad\sup_{0<h\leq h_m}\varepsilon_m(h)\leq m^{-2}. 0 < h m < min { 2 h m − 1 , 2 h ∗ , m 2 1 , 2 c + + m 2 d J } , 0 < h ≤ h m sup ε m ( h ) ≤ m − 2 .
Set L h : = m L_h:=m L h := m for h ∈ ( h m + 1 , h m ] h\in(h_{m+1},h_m] h ∈ ( h m + 1 , h m ] and assign it any positive value for larger h h h . Then
L h → h → 0 ∞ , h L h ≤ L h − 1 → h → 0 0. L_h\xrightarrow[h\to0]{}\infty,\qquad hL_h\leq L_h^{-1}\xrightarrow[h\to0]{}0. L h h → 0 ∞ , h L h ≤ L h − 1 h → 0 0.
Moreover,
sup t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ L h 2 ∥ X u t , h , c − B u t , h , c ∥ L 2 ≤ L h − 2 . \sup_{\substack{t\in J,\ c\in[c_-,c_+]\\0\leq u\leq L_h^2}}\left\|X_u^{t,h,c}-B_u^{t,h,c}\right\|_{L^2}\leq L_h^{-2}. t ∈ J , c ∈ [ c − , c + ] 0 ≤ u ≤ L h 2 sup X u t , h , c − B u t , h , c L 2 ≤ L h − 2 .
In particular, this controls [ 0 , 2 L h ] [0,2L_h] [ 0 , 2 L h ] for large L h L_h L h . Tonelli’s theorem and Cauchy–Schwarz therefore give
E [ ∫ 0 L h ∣ X u t , h , c − B u t , h , c ∣ 2 d u ] ≤ L h ε L h ( h ) 2 ≤ L h − 3 , \mathbb{E}\left[\int_0^{L_h}\left|X_u^{t,h,c}-B_u^{t,h,c}\right|^2\,\mathrm{d}u\right]\leq L_h\varepsilon_{L_h}(h)^2\leq L_h^{-3}, E [ ∫ 0 L h X u t , h , c − B u t , h , c 2 d u ] ≤ L h ε L h ( h ) 2 ≤ L h − 3 ,
and
E [ ∣ X L h t , h , c − B L h t , h , c ∣ ] ≤ ε L h ( h ) ≤ L h − 2 . \mathbb{E}\left[\left|X_{L_h}^{t,h,c}-B_{L_h}^{t,h,c}\right|\right]\leq\varepsilon L_h(h)\leq L_h^{-2}. E [ X L h t , h , c − B L h t , h , c ] ≤ ε L h ( h ) ≤ L h − 2 .
(FT1) and lemma 6.4 give
U L h ( X t , h , c ; ( F t + c h u ) u ≤ L h ) = U L h ( X t , h , c ; G ‾ t , h , c ) , \mathcal{U}_{L_h}\left(X^{t,h,c};(\mathcal{F}_{t+chu})_{u\leq L_h}\right)=\mathcal{U}_{L_h}\left(X^{t,h,c};\overline{G}^{t,h,c}\right), U L h ( X t , h , c ; ( F t + c h u ) u ≤ L h ) = U L h ( X t , h , c ; G t , h , c ) ,
and
U L h ( B t , h , c ; G ‾ t , h , c ) = u H p a s t ( L h ) . \mathcal{U}_{L_h}\left(B^{t,h,c};\overline{G}^{t,h,c}\right)=u_H^{\mathrm{past}}(L_h). U L h ( B t , h , c ; G t , h , c ) = u H past ( L h ) .
For 0 ≤ v ≤ 1 0\leq v\leq1 0 ≤ v ≤ 1 and j = 0 , … , L h − 1 j=0,\ldots,L_h-1 j = 0 , … , L h − 1 , (FT3) and strict positivity of ρ \rho ρ give
E [ ∣ X j + v t , h , c − X j t , h , c ∣ 2 ] = E [ ∣ Θ t + c h ( j + v ) − Θ t + c h j ∣ 2 ] ρ ( t ) 2 ( c h ) 2 H ≤ C ( c h v ) 2 H ρ ( t ) 2 ( c h ) 2 H ≤ C J v 2 H . \mathbb{E}\left[\left|X_{j+v}^{t,h,c}-X_j^{t,h,c}\right|^2\right]=\frac{\mathbb{E}\left[\left|\Theta_{t+ch(j+v)}-\Theta_{t+chj}\right|^2\right]}{\rho(t)^2(ch)^{2H}}\leq\frac{C(chv)^{2H}}{\rho(t)^2(ch)^{2H}}\leq C_Jv^{2H}. E [ X j + v t , h , c − X j t , h , c 2 ] = ρ ( t ) 2 ( c h ) 2 H E [ Θ t + c h ( j + v ) − Θ t + c hj 2 ] ≤ ρ ( t ) 2 ( c h ) 2 H C ( c h v ) 2 H ≤ C J v 2 H .
A rule with time step one therefore bounds both reset values by C ( 1 + L h ) C(1+L_h) C ( 1 + L h ) , uniformly in t t t , c c c , h h h . The stability estimate yields
∣ U L h ( X t , h , c ) − u H p a s t ( L h ) ∣ L h ≤ L h − 3 + 1 2 L h − 4 + C L h − 2 → h → 0 0. (115) \frac{\left|\mathcal{U}_{L_h}\left(X^{t,h,c}\right)-u_H^{\mathrm{past}}(L_h)\right|}{L_h}\leq L_h^{-3}+\frac{1}{2}L_h^{-4}+CL_h^{-2}\xrightarrow[h\to0]{}0. \tag*{(115)} L h U L h ( X t , h , c ) − u H past ( L h ) ≤ L h − 3 + 2 1 L h − 4 + C L h − 2 h → 0 0. (115)
uniformly. Now use u H p a s t ( L h ) / L h → h → 0 κ H u_H^{\mathrm{past}}(L_h)/L_h\xrightarrow[h\to0]{}\kappa_H u H past ( L h ) / L h h → 0 κ H . Let R h : = q L h + r h R_h:=qL_h+r_h R h := q L h + r h , with q q q in a compact subset of ( 0 , ∞ ) (0,\infty) ( 0 , ∞ ) and r h = O ( 1 ) r_h=O(1) r h = O ( 1 ) uniformly in all parameters. For small h h h , 0 < R h ≤ L h 2 0<R_h\leq L_h^2 0 < R h ≤ L h 2 and R h ≍ L h R_h\asymp L_h R h ≍ L h . The same coupling gives, uniformly,
E [ ∫ 0 R h ∣ X u t , h , c − B u t , h , c ∣ 2 d u ] ≤ R h L h − 4 = o ( R h ) , \mathbb{E}\left[\int_0^{R_h}\left|X_u^{t,h,c}-B_u^{t,h,c}\right|^2\,du\right]\leq R_hL_h^{-4}=o(R_h), E [ ∫ 0 R h X u t , h , c − B u t , h , c 2 d u ] ≤ R h L h − 4 = o ( R h ) ,
and
E [ ∣ X R h t , h , c − B R h t , h , c ∣ ] ≤ L h − 2 = o ( R h ) . \mathbb{E}\left[\left|X_{R_h}^{t,h,c}-B_{R_h}^{t,h,c}\right|\right]\leq L_h^{-2}=o(R_h). E [ X R h t , h , c − B R h t , h , c ] ≤ L h − 2 = o ( R h ) .
Fix 0 < q − < q + < ∞ 0<q_-<q_+<\infty 0 < q − < q + < ∞ . The values obtained with time step one are O ( R h ) O(R_h) O ( R h ) uniformly in q ∈ [ q − , q + ] q\in[q_-,q_+] q ∈ [ q − , q + ] , while
sup q ∈ [ q − , q + ] ∣ u H p a s t ( q L h + r h ) q L h + r h − κ H ∣ → h → 0 0 , \sup_{q\in[q_-,q_+]}\left|\frac{u_H^{\mathrm{past}}(qL_h+r_h)}{qL_h+r_h}-\kappa_H\right|\xrightarrow[h\to0]{}0, q ∈ [ q − , q + ] sup q L h + r h u H past ( q L h + r h ) − κ H h → 0 0 ,
since the horizons diverge uniformly. Stability then gives the transfer statement with L h L_h L h replaced by R h R_h R h . □ \square □
Block estimates
Lemma H.1 (block reset decomposition). Let θ ∈ C ( [ 0 , T ] ; R ) \theta\in\mathcal{C}([0,T];\mathbb{R}) θ ∈ C ([ 0 , T ] ; R ) and let ϕ \phi ϕ be right-continuous with finite variation. For 0 ≤ r 0 < r 1 < ⋯ < r N ≤ T 0\leq r_0<r_1<\cdots<r_N\leq T 0 ≤ r 0 < r 1 < ⋯ < r N ≤ T , the reset variations on the blocks satisfy
∑ i = 1 N [ ∣ ϕ r i − 1 − θ r i − 1 ∣ + TV ( r i − 1 , r i ] ( ϕ ) + ∣ ϕ r i − θ r i ∣ ] ≤ TV ( r 0 , r N ] ( ϕ ) + 2 ∑ j = 0 N ∣ ϕ r j − θ r j ∣ . (116) \sum_{i=1}^{N}\left[|\phi_{r_{i-1}}-\theta_{r_{i-1}}|+\operatorname{TV}_{(r_{i-1},r_i]}(\phi)+|\phi_{r_i}-\theta_{r_i}|\right]\leq\operatorname{TV}_{(r_0,r_N]}(\phi)+2\sum_{j=0}^{N}|\phi_{r_j}-\theta_{r_j}|. \tag*{(116)} i = 1 ∑ N [ ∣ ϕ r i − 1 − θ r i − 1 ∣ + TV ( r i − 1 , r i ] ( ϕ ) + ∣ ϕ r i − θ r i ∣ ] ≤ TV ( r 0 , r N ] ( ϕ ) + 2 j = 0 ∑ N ∣ ϕ r j − θ r j ∣. (116) If b : [ 0 , T ] → ( 0 , ∞ ) b:[0,T]\to(0,\infty) b : [ 0 , T ] → ( 0 , ∞ ) is Borel and b ≤ b + b\leq b_+ b ≤ b + , then the weighted reset variations satisfy
∑ i = 1 N [ b r i − 1 ∣ ϕ r i − 1 − θ r i − 1 ∣ + ∫ ( r i − 1 , r i ] b t d ∣ ϕ ∣ t + b r i ∣ ϕ r i − θ r i ∣ ] ≤ ∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + 2 b + ∑ j = 0 N ∣ ϕ r j − θ r j ∣ . (117) \sum_{i=1}^{N}\left[b_{r_{i-1}}|\phi_{r_{i-1}}-\theta_{r_{i-1}}|+\int_{(r_{i-1},r_i]}b_t\,d|\phi|_t+b_{r_i}|\phi_{r_i}-\theta_{r_i}|\right]\leq\int_{(r_0,r_N]}b_t\,d|\phi|_t+2b_+\sum_{j=0}^{N}|\phi_{r_j}-\theta_{r_j}|. \tag*{(117)} i = 1 ∑ N [ b r i − 1 ∣ ϕ r i − 1 − θ r i − 1 ∣ + ∫ ( r i − 1 , r i ] b t d ∣ ϕ ∣ t + b r i ∣ ϕ r i − θ r i ∣ ] ≤ ∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + 2 b + j = 0 ∑ N ∣ ϕ r j − θ r j ∣. (117) Proof. Additivity on the disjoint intervals ( r i − 1 , r i ] (r_{i-1},r_i] ( r i − 1 , r i ] makes the left-hand side of (116) equal to
TV ( r 0 , r N ] ( ϕ ) + ∣ ϕ r 0 − θ r 0 ∣ + 2 ∑ j = 1 N − 1 ∣ ϕ r j − θ r j ∣ + ∣ ϕ r N − θ r N ∣ \operatorname{TV}_{(r_0,r_N]}(\phi)+|\phi_{r_0}-\theta_{r_0}|+2\sum_{j=1}^{N-1}|\phi_{r_j}-\theta_{r_j}|+|\phi_{r_N}-\theta_{r_N}| TV ( r 0 , r N ] ( ϕ ) + ∣ ϕ r 0 − θ r 0 ∣ + 2 j = 1 ∑ N − 1 ∣ ϕ r j − θ r j ∣ + ∣ ϕ r N − θ r N ∣ ≤ TV ( r 0 , r N ] ( ϕ ) + 2 ∑ j = 0 N ∣ ϕ r j − θ r j ∣ . \leq\operatorname{TV}_{(r_0,r_N]}(\phi)+2\sum_{j=0}^{N}|\phi_{r_j}-\theta_{r_j}|. ≤ TV ( r 0 , r N ] ( ϕ ) + 2 j = 0 ∑ N ∣ ϕ r j − θ r j ∣.
Likewise, additivity gives the first expression below for the weighted left-hand side. The bound then follows from b ≤ b + b \le b_{+} b ≤ b + :
∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + b r 0 ∣ ϕ r 0 − θ r 0 ∣ + 2 ∑ j = 1 N − 1 b r j ∣ ϕ r j − θ r j ∣ + b r N ∣ ϕ r N − θ r N ∣ ≤ ∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + 2 b + ∑ j = 0 N ∣ ϕ r j − θ r j ∣ . \begin{aligned}
\int_{(r_0,r_N]} b_t\,\mathrm{d}|\phi|_t+b_{r_0}|\phi_{r_0}-\theta_{r_0}|+2\sum_{j=1}^{N-1}b_{r_j}|\phi_{r_j}-\theta_{r_j}|+b_{r_N}|\phi_{r_N}-\theta_{r_N}| \\
&\le\int_{(r_0,r_N]} b_t\,\mathrm{d}|\phi|_t+2b_{+}\sum_{j=0}^{N}|\phi_{r_j}-\theta_{r_j}|.
\end{aligned} ∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + b r 0 ∣ ϕ r 0 − θ r 0 ∣ + 2 j = 1 ∑ N − 1 b r j ∣ ϕ r j − θ r j ∣ + b r N ∣ ϕ r N − θ r N ∣ ≤ ∫ ( r 0 , r N ] b t d ∣ ϕ ∣ t + 2 b + j = 0 ∑ N ∣ ϕ r j − θ r j ∣.
Lemma H.2 (separator points). Let e e e be a measurable process with E [ ∫ 0 T e t 2 d t ] < ∞ \mathbb{E}\left[\int_0^T e_t^2\,\mathrm{d}t\right]<\infty E [ ∫ 0 T e t 2 d t ] < ∞ . Let J 1 , … , J N J_1,\ldots,J_N J 1 , … , J N be disjoint intervals of the same length h > 0 h>0 h > 0 . There are deterministic points r i ∈ J i r_i\in J_i r i ∈ J i such that
E [ ∣ e r i ∣ 2 ] ≤ 1 h ∫ J i E [ ∣ e t ∣ 2 ] d t . (118) \mathbb{E}[|e_{r_i}|^2]\le\frac{1}{h}\int_{J_i}\mathbb{E}[|e_t|^2]\,\mathrm{d}t. \tag*{(118)} E [ ∣ e r i ∣ 2 ] ≤ h 1 ∫ J i E [ ∣ e t ∣ 2 ] d t . (118) Then,
∑ i = 1 N E [ ∣ e r i ∣ ] ≤ ( N h E [ ∫ ⋃ i J i e t 2 d t ] ) 1 / 2 . (119) \sum_{i=1}^{N}\mathbb{E}[|e_{r_i}|]\le\left(\frac{N}{h}\mathbb{E}\left[\int_{\bigcup_iJ_i}e_t^2\,\mathrm{d}t\right]\right)^{1/2}. \tag*{(119)} i = 1 ∑ N E [ ∣ e r i ∣ ] ≤ ( h N E [ ∫ ⋃ i J i e t 2 d t ] ) 1/2 . (119) Proof. Choose r i ∈ J i r_i\in J_i r i ∈ J i with E [ ∣ e r i ∣ 2 ] ≤ h − 1 ∫ J i E [ ∣ e t ∣ 2 ] d t \mathbb{E}[|e_{r_i}|^2]\le h^{-1}\int_{J_i}\mathbb{E}[|e_t|^2]\,\mathrm{d}t E [ ∣ e r i ∣ 2 ] ≤ h − 1 ∫ J i E [ ∣ e t ∣ 2 ] d t . Such a point exists by the integral average. Two applications of Cauchy–Schwarz give
∑ i = 1 N E [ ∣ e r i ∣ ] ≤ ∑ i = 1 N ( E [ ∣ e r i ∣ 2 ] ) 1 / 2 \sum_{i=1}^{N}\mathbb{E}[|e_{r_i}|]\le\sum_{i=1}^{N}\left(\mathbb{E}[|e_{r_i}|^2]\right)^{1/2} i = 1 ∑ N E [ ∣ e r i ∣ ] ≤ i = 1 ∑ N ( E [ ∣ e r i ∣ 2 ] ) 1/2 ≤ ( N ∑ i = 1 N E [ ∣ e r i ∣ 2 ] ) 1 / 2 \le\left(N\sum_{i=1}^{N}\mathbb{E}[|e_{r_i}|^2]\right)^{1/2} ≤ ( N i = 1 ∑ N E [ ∣ e r i ∣ 2 ] ) 1/2 ≤ ( N h ∑ i = 1 N ∫ J i E [ ∣ e t ∣ 2 ] d t ) 1 / 2 \le\left(\frac{N}{h}\sum_{i=1}^{N}\int_{J_i}\mathbb{E}[|e_t|^2]\,\mathrm{d}t\right)^{1/2} ≤ ( h N i = 1 ∑ N ∫ J i E [ ∣ e t ∣ 2 ] d t ) 1/2 = ( N h E [ ∫ ⋃ i J i e t 2 d t ] ) 1 / 2 . =\left(\frac{N}{h}\mathbb{E}\left[\int_{\bigcup_iJ_i}e_t^2\,\mathrm{d}t\right]\right)^{1/2}. = ( h N E [ ∫ ⋃ i J i e t 2 d t ] ) 1/2 .
Lemma H.3 (short-interval control). Assume (38) , and let a , b a,b a , b be bounded above. Put h λ : = λ 1 / ( 1 + H ) h_\lambda:=\lambda^{1/(1+H)} h λ := λ 1/ ( 1 + H ) and α H : = 2 H / ( 1 + H ) \alpha_H:=2H/(1+H) α H := 2 H / ( 1 + H ) . On every interval I = [ s , s + r ] ⊂ [ 0 , T ] I=[s,s+r]\subset[0,T] I = [ s , s + r ] ⊂ [ 0 , T ] , there is a control that starts from Θ s \Theta_s Θ s , resets to Θ s + r \Theta_{s+r} Θ s + r at the end, and has expected cost at most
C λ α H ( r + h λ ) , (120) C\lambda^{\alpha_H}(r+h_\lambda), \tag*{(120)} C λ α H ( r + h λ ) , (120) where C C C is independent of s , r , λ s,r,\lambda s , r , λ , for all sufficiently small λ > 0 \lambda>0 λ > 0 .
Proof. Take h λ < h 0 h_\lambda<h_0 h λ < h 0 . For r > 0 r>0 r > 0 , set N : = ⌈ r / h λ ⌉ N:=\lceil r/h_\lambda\rceil N := ⌈ r / h λ ⌉ and t k : = s + min ( k h λ , r ) t_k:=s+\min(kh_\lambda,r) t k := s + min ( k h λ , r ) , 0 ≤ k ≤ N 0\le k\le N 0 ≤ k ≤ N . Use ϕ t : = Θ t k \phi_t:=\Theta_{t_k} ϕ t := Θ t k on [ t k , t k + 1 ) [t_k,t_{k+1}) [ t k , t k + 1 ) , with ϕ s + r : = Θ s + r \phi_{s+r}:=\Theta_{s+r} ϕ s + r := Θ s + r . The initial adjustment is zero, and the increment bound controls the tracking cost. For the variation cost, Cauchy–Schwarz and N ≤ r / h λ + 1 N\le r/h_\lambda+1 N ≤ r / h λ + 1 give the remaining bounds:
E [ ∫ I ∣ ϕ t − Θ t ∣ 2 d t ] ≤ C ∑ k = 0 N − 1 ∫ 0 t k + 1 − t k u 2 H d u ≤ C r h λ 2 H , \mathbb{E}\left[\int_I|\phi_t-\Theta_t|^2\,\mathrm{d}t\right]\le C\sum_{k=0}^{N-1}\int_0^{t_{k+1}-t_k}u^{2H}\,\mathrm{d}u\le Crh_\lambda^{2H}, E [ ∫ I ∣ ϕ t − Θ t ∣ 2 d t ] ≤ C k = 0 ∑ N − 1 ∫ 0 t k + 1 − t k u 2 H d u ≤ C r h λ 2 H , and
E [ TV [ s , s + r ] a u g ( ϕ ) ] = ∑ k = 0 N − 1 E [ ∣ Θ t k + 1 − Θ t k ∣ ] ≤ C N h λ H ≤ C ( r h λ H − 1 + h λ H ) . \mathbb{E}\left[\operatorname{TV}^{\mathrm{aug}}_{[s,s+r]}(\phi)\right]=\sum_{k=0}^{N-1}\mathbb{E}[|\Theta_{t_{k+1}}-\Theta_{t_k}|]\le CNh_\lambda^H\le C(rh_\lambda^{H-1}+h_\lambda^H). E [ TV [ s , s + r ] aug ( ϕ ) ] = k = 0 ∑ N − 1 E [ ∣ Θ t k + 1 − Θ t k ∣ ] ≤ C N h λ H ≤ C ( r h λ H − 1 + h λ H ) . Since λ = h λ 1 + H \lambda=h_\lambda^{1+H} λ = h λ 1 + H and λ α H = h λ 2 H \lambda^{\alpha_H}=h_\lambda^{2H} λ α H = h λ 2 H , the weighted cost is at most
C { r h λ 2 H + λ r h λ H − 1 + λ h λ H } ≤ C λ α H ( r + h λ ) . C\{rh_\lambda^{2H}+\lambda rh_\lambda^{H-1}+\lambda h_\lambda^H\}\le C\lambda^{\alpha_H}(r+h_\lambda). C { r h λ 2 H + λ r h λ H − 1 + λ h λ H } ≤ C λ α H ( r + h λ ) . The case r = 0 r=0 r = 0 is immediate.
The Gaussian Volterra tangent
Proof of proposition 6.8. Fix a compact interval J ⊂ ( 0 , T ) J \subset(0,T) J ⊂ ( 0 , T ) , 0 < c − < c + < ∞ 0 < c_{-} < c_{+} < \infty 0 < c − < c + < ∞ , and R < ∞ R < \infty R < ∞ . The limits below are uniform in t ∈ J t \in J t ∈ J , c ∈ [ c − , c + ] c \in[c_{-},c_{+}] c ∈ [ c − , c + ] , and u ∈ [ 0 , R ] u \in[0,R] u ∈ [ 0 , R ] .
We first construct the tangent and its observation filtration. On an extension ( Ω ‾ , F ‾ , P ‾ ) (\overline{\Omega},\overline{\mathcal{F}},\overline{\mathbb{P}}) ( Ω , F , P ) , extend W W W to a two-sided Brownian motion W ~ \widetilde{W} W by an independent past and, if needed, an independent continuation after T T T . Set β : = H − 1 / 2 ∈ ( − 1 / 2 , 1 / 2 ) \beta:= H - 1/2 \in(-1/2,1/2) β := H − 1/2 ∈ ( − 1/2 , 1/2 ) and
W ~ t , h : R ⟶ L 2 ( Ω ‾ , P ‾ ) \widetilde{W}^{t,h}\colon\mathbb{R} \longrightarrow L^{2}(\overline{\Omega},\overline{\mathbb{P}}) W t , h : R ⟶ L 2 ( Ω , P )
z ⟼ h − 1 / 2 ( W ~ t + h z − W ~ t ) . z \longmapsto h^{-1/2}\left(\widetilde{W}_{t+hz}-\widetilde{W}_{t}\right). z ⟼ h − 1/2 ( W t + h z − W t ) .
This is two-sided standard Brownian motion. Use the kernel g u H g^{H}_{u} g u H from (64) . For general H H H , g u H ∈ L 2 ( R ) g^{H}_{u} \in L^{2}(\mathbb{R}) g u H ∈ L 2 ( R ) : its local squared singularities have exponent 2 β = 2 H − 1 > − 1 2\beta= 2H - 1 > -1 2 β = 2 H − 1 > − 1 , while
g u H ( z ) = O u ( ∣ z ∣ H − 3 / 2 ) ( z → − ∞ ) , 2 H − 3 < − 1. g^{H}_{u}(z) = O_{u}\left(|z|^{H-3/2}\right) \quad(z \to-\infty), \qquad2H - 3 < -1. g u H ( z ) = O u ( ∣ z ∣ H − 3/2 ) ( z → − ∞ ) , 2 H − 3 < − 1.
Wiener integration is the linear isometry from L 2 ( R ) L^{2}(\mathbb{R}) L 2 ( R ) into L 2 ( Ω ‾ , P ‾ ) L^{2}(\overline{\Omega},\overline{\mathbb{P}}) L 2 ( Ω , P ) satisfying
E [ ( ∫ R f d W ~ ) ( ∫ R g d W ~ ) ] = ∫ R f ( z ) g ( z ) d z , f , g ∈ L 2 ( R ) . \mathbb{E}\left[\left(\int_{\mathbb{R}} f\,d\widetilde{W}\right)\left(\int_{\mathbb{R}} g\,d\widetilde{W}\right)\right] = \int_{\mathbb{R}} f(z)g(z)\,dz,\qquad f,g \in L^{2}(\mathbb{R}). E [ ( ∫ R f d W ) ( ∫ R g d W ) ] = ∫ R f ( z ) g ( z ) d z , f , g ∈ L 2 ( R ) .
The kernel identities
g a u H ( a z ) = a H − 1 / 2 g u H ( z ) ( a > 0 ) , g^{H}_{au}(az) = a^{H-1/2}g^{H}_{u}(z)\qquad(a>0), g a u H ( a z ) = a H − 1/2 g u H ( z ) ( a > 0 ) ,
and
g u H ( z ) − g v H ( z ) = g u − v H ( z − v ) g^{H}_{u}(z)-g^{H}_{v}(z)=g^{H}_{u-v}(z-v) g u H ( z ) − g v H ( z ) = g u − v H ( z − v )
give, by scaling, translation, and g − u H ( z ) = − g u H ( z + u ) g^{H}_{-u}(z)=-g^{H}_{u}(z+u) g − u H ( z ) = − g u H ( z + u ) ,
∥ g 1 H ∥ 2 2 = ∫ 0 1 ( 1 − z ) 2 H − 1 d z + ∫ 0 ∞ ( ( 1 + z ) H − 1 / 2 − z H − 1 / 2 ) 2 d z = Γ H . \lVert g^{H}_{1}\rVert_{2}^{2}=\int_{0}^{1}(1-z)^{2H-1}\,dz+\int_{0}^{\infty}\left((1+z)^{H-1/2}-z^{H-1/2}\right)^{2}\,dz=\Gamma_{H}. ∥ g 1 H ∥ 2 2 = ∫ 0 1 ( 1 − z ) 2 H − 1 d z + ∫ 0 ∞ ( ( 1 + z ) H − 1/2 − z H − 1/2 ) 2 d z = Γ H .
Then,
∥ g u H ∥ 2 2 = Γ H ∣ u ∣ 2 H , \lVert g^{H}_{u}\rVert_{2}^{2}=\Gamma_{H}|u|^{2H}, ∥ g u H ∥ 2 2 = Γ H ∣ u ∣ 2 H ,
and
∥ g u H − g v H ∥ 2 2 = Γ H ∣ u − v ∣ 2 H . \lVert g^{H}_{u}-g^{H}_{v}\rVert_{2}^{2}=\Gamma_{H}|u-v|^{2H}. ∥ g u H − g v H ∥ 2 2 = Γ H ∣ u − v ∣ 2 H .
Polarization therefore identifies the covariance:
⟨ g u H , g v H ⟩ L 2 ( R ) = Γ H 2 ( ∣ u ∣ 2 H + ∣ v ∣ 2 H − ∣ u − v ∣ 2 H ) = Γ H R H ( u , v ) . \left\langle g^{H}_{u},g^{H}_{v}\right\rangle_{L^{2}(\mathbb{R})}=\frac{\Gamma_{H}}{2}\left(|u|^{2H}+|v|^{2H}-|u-v|^{2H}\right)=\Gamma_{H}R_{H}(u,v). ⟨ g u H , g v H ⟩ L 2 ( R ) = 2 Γ H ( ∣ u ∣ 2 H + ∣ v ∣ 2 H − ∣ u − v ∣ 2 H ) = Γ H R H ( u , v ) .
Consequently,
B t , h : { R ⟶ L 2 ( Ω ‾ , P ‾ ) , u ⟼ sign ( c K q ( t ) ) Γ H ∫ R g u H ( z ) d W ~ z t , h , (121) B^{t,h}\colon
\begin{cases}
\mathbb{R}\longrightarrow L^{2}(\overline{\Omega},\overline{\mathbb{P}}),\\
u\longmapsto\frac{\operatorname{sign}(cKq(t))}{\sqrt{\Gamma_{H}}}\int_{\mathbb{R}}g^{H}_{u}(z)\,d\widetilde{W}^{t,h}_{z},
\end{cases} \tag*{(121)} B t , h : { R ⟶ L 2 ( Ω , P ) , u ⟼ Γ H sign ( cK q ( t )) ∫ R g u H ( z ) d W z t , h , (121)
has the standard two-sided fBm law, using its continuous Gaussian version. Let U U U be the added Brownian path on ( − ∞ , 0 ] (-\infty,0] ( − ∞ , 0 ] , an independent C l o c ( ( − ∞ , 0 ] ; R ) \mathcal{C}_{\mathrm{loc}}((-\infty,0];\mathbb{R}) C loc (( − ∞ , 0 ] ; R ) -valued variable. For u ≥ 0 u \ge0 u ≥ 0 with t + h u ≤ T t+hu \le T t + h u ≤ T ,
F t + h u W ∨ σ ( U ) = σ ( W ~ z t , h : z ≤ u ) = σ ( B v t , h : v ≤ u ) . \mathcal{F}^{W}_{t+hu}\vee\sigma(U)=\sigma\left(\widetilde{W}^{t,h}_{z}:z\le u\right)=\sigma\left(B^{t,h}_{v}:v\le u\right). F t + h u W ∨ σ ( U ) = σ ( W z t , h : z ≤ u ) = σ ( B v t , h : v ≤ u ) .
The first equality follows from W 0 = 0 W_{0}=0 W 0 = 0 : W ~ − t / h t , h = − h − 1 / 2 W t \widetilde{W}^{t,h}_{-t/h}=-h^{-1/2}W_{t} W − t / h t , h = − h − 1/2 W t , so the local past recovers W t W_{t} W t and then the original Brownian path. The second equality is lemma A.1. Taking B t , h , c : = B t , c h B^{t,h,c}:=B^{t,ch} B t , h , c := B t , c h and U t , h , c : = U U^{t,h,c}:=U U t , h , c := U proves (FT1). To establish the uniform L 2 L^{2} L 2 approximation, split the increment at t t t and substitute s = t + h z s=t+hz s = t + h z :
h − H { Θ t + h u − Θ t − F ( t + h u ) + F ( t ) } = ∫ R g h , t , u ( z ) d W ~ z t , h , (122) h^{-H}\{\Theta_{t+hu}-\Theta_{t}-F(t+hu)+F(t)\}=\int_{\mathbb{R}}g_{h,t,u}(z)\,d\widetilde{W}^{t,h}_{z}, \tag*{(122)} h − H { Θ t + h u − Θ t − F ( t + h u ) + F ( t )} = ∫ R g h , t , u ( z ) d W z t , h , (122)
where g h , t , u ∈ L 2 ( R ) g_{h,t,u}\in L^{2}(\mathbb{R}) g h , t , u ∈ L 2 ( R ) is
g h , t , u : { R ⟶ R , z ⟼ h − β { [ K ( h ( u − z ) ) − K ( − h z ) ] q ( t + h z ) , − t / h ≤ z < 0 , K ( h ( u − z ) ) q ( t + h z ) , 0 ≤ z ≤ u , 0 , otherwise . g_{h,t,u}\colon
\begin{cases}
\mathbb{R}\longrightarrow\mathbb{R},\\
z\longmapsto h^{-\beta}
\begin{cases}
[K(h(u-z))-K(-hz)]q(t+hz),& -t/h\le z<0,\\
K(h(u-z))q(t+hz),& 0\le z\le u,\\
0,& \text{otherwise}.
\end{cases}
\end{cases} g h , t , u : ⎩ ⎨ ⎧ R ⟶ R , z ⟼ h − β ⎩ ⎨ ⎧ [ K ( h ( u − z )) − K ( − h z )] q ( t + h z ) , K ( h ( u − z )) q ( t + h z ) , 0 , − t / h ≤ z < 0 , 0 ≤ z ≤ u , otherwise .
Endpoint kernel values may be chosen arbitrarily. We prove
sup t ∈ J sup 0 ≤ u ≤ R ∥ g h , t , u − c K q ( t ) g u H ∥ L 2 ( R ) ⟶ 0. (123) \sup_{t\in J}\sup_{0\le u\le R}\|g_{h,t,u}-c_Kq(t)g_u^H\|_{L^2(\mathbb{R})}\longrightarrow0. \tag*{(123)} t ∈ J sup 0 ≤ u ≤ R sup ∥ g h , t , u − c K q ( t ) g u H ∥ L 2 ( R ) ⟶ 0. (123)
Fix A > 1 + R A>1+R A > 1 + R and take h A < inf J hA<\inf J h A < inf J . Define the relative kernel error
ζ K : { ( 0 , T ] ⟶ [ 0 , ∞ ) v ⟼ sup 0 < r ≤ v ∣ K ( r ) c K r β − 1 ∣ . \zeta_K:\begin{cases}
(0,T]\longrightarrow[0,\infty)\\
v\longmapsto\sup_{0<r\le v}\left|\frac{K(r)}{c_Kr^\beta}-1\right|.
\end{cases} ζ K : { ( 0 , T ] ⟶ [ 0 , ∞ ) v ⟼ sup 0 < r ≤ v c K r β K ( r ) − 1 .
By (61) , ζ K ( v ) → 0 \zeta_K(v)\to0 ζ K ( v ) → 0 as v → 0 v\to0 v → 0 . For z ∈ [ − A , u ] z\in[-A,u] z ∈ [ − A , u ] and positive arguments x ≤ A + R x\le A+R x ≤ A + R ,
∣ h − β K ( h x ) q ( t + h z ) − c K q ( t ) x β ∣ ≤ ∣ c K ∣ x β ( ζ K ( h ( A + R ) ) + ω q ( h A ) ) . \left|h^{-\beta}K(hx)q(t+hz)-c_Kq(t)x^\beta\right|
\le|c_K|x^\beta\bigl(\zeta_K(h(A+R))+\omega_q(hA)\bigr). h − β K ( h x ) q ( t + h z ) − c K q ( t ) x β ≤ ∣ c K ∣ x β ( ζ K ( h ( A + R )) + ω q ( h A ) ) .
Apply this to x = u − z x=u-z x = u − z and, on z < 0 z<0 z < 0 , to x = − z x=-z x = − z . Consequently,
∫ − A R ∣ g h , t , u ( z ) − c K q ( t ) g u H ( z ) ∣ 2 d z ≤ C ( ζ K ( h ( A + R ) ) + ω q ( h A ) ) 2 × ( ∫ − A u ( u − z ) 2 β d z + ∫ − A 0 ( − z ) 2 β d z ) = C 2 H ( ζ K ( h ( A + R ) ) + ω q ( h A ) ) 2 ( ( A + u ) 2 H + A 2 H ) ≤ C A , R ( ζ K ( h ( A + R ) ) + ω q ( h A ) ) 2 ⟶ 0. \begin{aligned}
\int_{-A}^{R}|g_{h,t,u}(z)-c_Kq(t)g_u^H(z)|^2\,\mathrm{d}z
&\le C\bigl(\zeta_K(h(A+R))+\omega_q(hA)\bigr)^2\\
&\quad\times\left(\int_{-A}^{u}(u-z)^{2\beta}\,\mathrm{d}z+\int_{-A}^{0}(-z)^{2\beta}\,\mathrm{d}z\right)\\
&=\frac{C}{2H}\bigl(\zeta_K(h(A+R))+\omega_q(hA)\bigr)^2
\bigl((A+u)^{2H}+A^{2H}\bigr)\\
&\le C_{A,R}\bigl(\zeta_K(h(A+R))+\omega_q(hA)\bigr)^2
\longrightarrow0.
\end{aligned} ∫ − A R ∣ g h , t , u ( z ) − c K q ( t ) g u H ( z ) ∣ 2 d z ≤ C ( ζ K ( h ( A + R )) + ω q ( h A ) ) 2 × ( ∫ − A u ( u − z ) 2 β d z + ∫ − A 0 ( − z ) 2 β d z ) = 2 H C ( ζ K ( h ( A + R )) + ω q ( h A ) ) 2 ( ( A + u ) 2 H + A 2 H ) ≤ C A , R ( ζ K ( h ( A + R )) + ω q ( h A ) ) 2 ⟶ 0.
For the remaining tail − t / h ≤ z < − A -t/h\le z<-A − t / h ≤ z < − A , the mean-value theorem and the derivative bound give
h − β ∣ K ( h ( u − z ) ) − K ( − h z ) ∣ ≤ h − β ( h u ) sup 0 ≤ θ ≤ u ∣ K ′ ( h ( − z + θ ) ) ∣ ≤ C R h − β + 1 + H − 3 / 2 ∣ z ∣ H − 3 / 2 = C R ∣ z ∣ H − 3 / 2 . (124) \begin{aligned}
h^{-\beta}|K(h(u-z))-K(-hz)|
\le h^{-\beta}(hu)\sup_{0\le\theta\le u}|K'(h(-z+\theta))| \\
&\le C_Rh^{-\beta+1+H-3/2}|z|^{H-3/2}
=C_R|z|^{H-3/2}. \tag*{(124)}
\end{aligned} h − β ∣ K ( h ( u − z )) − K ( − h z ) ∣ ≤ h − β ( h u ) 0 ≤ θ ≤ u sup ∣ K ′ ( h ( − z + θ )) ∣ ≤ C R h − β + 1 + H − 3/2 ∣ z ∣ H − 3/2 = C R ∣ z ∣ H − 3/2 . (124)
The limiting kernel has the same tail bound. Therefore
sup t , u t ∈ J , u ≤ R ∫ − ∞ − A ∣ g h , t , u − c K q ( t ) g u H ∣ 2 d z ≤ C R ∫ A ∞ z 2 H − 3 d z = C R 2 − 2 H A 2 H − 2 ⟶ 0. \sup_{\substack{t,u\\ t\in J,\;u\le R}}\int_{-\infty}^{-A}|g_{h,t,u}-c_Kq(t)g_u^H|^2\,\mathrm{d}z
\le C_R\int_A^\infty z^{2H-3}\,\mathrm{d}z
=\frac{C_R}{2-2H}A^{2H-2}\longrightarrow0. t , u t ∈ J , u ≤ R sup ∫ − ∞ − A ∣ g h , t , u − c K q ( t ) g u H ∣ 2 d z ≤ C R ∫ A ∞ z 2 H − 3 d z = 2 − 2 H C R A 2 H − 2 ⟶ 0.
The omitted original past is also negligible:
sup t ∈ J , u ≤ R ∫ − ∞ − t / h ∣ g u H ( z ) ∣ 2 d z ≤ C R ( h inf J ) 2 − 2 H ⟶ 0. \sup_{\substack{t\in J,\;u\le R}}\int_{-\infty}^{-t/h}|g_u^H(z)|^2\,\mathrm{d}z
\le C_R\left(\frac{h}{\inf J}\right)^{2-2H}\longrightarrow0. t ∈ J , u ≤ R sup ∫ − ∞ − t / h ∣ g u H ( z ) ∣ 2 d z ≤ C R ( inf J h ) 2 − 2 H ⟶ 0.
First take h → 0 h\to0 h → 0 on the compact region, then A → ∞ A\to\infty A → ∞ . This proves (123) . All bounds remain uniform with h h h replaced by c h ch c h . The Wiener isometry and the normalization identity
c K q ( t ) ∫ R g u H ( z ) d W ~ z t , h = (118) ρ ( t ) B u t , h c_Kq(t)\int_{\mathbb{R}}g_u^H(z)\,\mathrm{d}\widetilde W_z^{t,h}
\overset{\text{(118)}}{=}\rho(t)B_u^{t,h} c K q ( t ) ∫ R g u H ( z ) d W z t , h = (118) ρ ( t ) B u t , h
handle the stochastic part. The drift satisfies
sup t ∈ J , u ≤ R ∣ F ( t + h u ) − F ( t ) ∣ h H ≤ [ F ] α F R α F h α F − H ⟶ 0. \sup_{\substack{t\in J,\;u\le R}}
\frac{|F(t+hu)-F(t)|}{h^H}
\le[F]_{\alpha_F}R^{\alpha_F}h^{\alpha_F-H}\longrightarrow0. t ∈ J , u ≤ R sup h H ∣ F ( t + h u ) − F ( t ) ∣ ≤ [ F ] α F R α F h α F − H ⟶ 0.
Since inf t ρ ( t ) > 0 \inf_t\rho(t)>0 inf t ρ ( t ) > 0 , division by ρ ( t ) \rho(t) ρ ( t ) proves (FT2), uniformly also in c c c . Finally, we verify the global increment bound and continuity. For 0 ≤ t ≤ T − h 0\le t\le T-h 0 ≤ t ≤ T − h , decompose
Θ t + h − Θ t = F ( t + h ) − F ( t ) + ∫ 0 t { K ( t + h − s ) − K ( t − s ) } q ( s ) d W s \Theta_{t+h}-\Theta_t
=F(t+h)-F(t)
+\int_0^t\{K(t+h-s)-K(t-s)\}q(s)\,\mathrm{d}W_s Θ t + h − Θ t = F ( t + h ) − F ( t ) + ∫ 0 t { K ( t + h − s ) − K ( t − s )} q ( s ) d W s
+ ∫ t t + h K ( t + h − s ) q ( s ) d W s . \qquad+\int_t^{t+h}K(t+h-s)q(s)\,\mathrm{d}W_s. + ∫ t t + h K ( t + h − s ) q ( s ) d W s .
The two stochastic integrals are orthogonal. On 0 < u < h 0<u<h 0 < u < h , bound the kernel difference by ∣ K ( u + h ) ∣ + ∣ K ( u ) ∣ |K(u+h)|+|K(u)| ∣ K ( u + h ) ∣ + ∣ K ( u ) ∣ . On u ≥ h u\ge h u ≥ h , use the mean-value theorem. Itô’s isometry and α F > H \alpha_F>H α F > H then yield, for small h ≤ 1 h\le1 h ≤ 1 ,
E [ ∣ Θ t + h − Θ t ∣ 2 ] ≤ C h 2 α F + C ∫ 0 T ∣ K ( u + h ) − K ( u ) ∣ 2 d u + C ∫ 0 h ∣ K ( u ) ∣ 2 d u ≤ C h 2 α F + C ∫ 0 2 h u 2 H − 1 d u + C h 2 ∫ h T u 2 H − 3 d u + C ∫ 0 h u 2 H − 1 d u ≤ C h 2 α F + C ( 2 h ) 2 H + h 2 H 2 H + C h 2 H − h 2 T 2 H − 2 2 − 2 H ≤ C ( h 2 α F + h 2 H ) ≤ C h 2 H . (125) \begin{aligned}
\mathbb{E}\left[|\Theta_{t+h}-\Theta_t|^2\right]\le Ch^{2\alpha_F}+C\int_0^T|K(u+h)-K(u)|^2\,\mathrm{d}u+C\int_0^h|K(u)|^2\,\mathrm{d}u \\
&\le Ch^{2\alpha_F}+C\int_0^{2h}u^{2H-1}\,\mathrm{d}u+Ch^2\int_h^T u^{2H-3}\,\mathrm{d}u+C\int_0^h u^{2H-1}\,\mathrm{d}u \\
&\le Ch^{2\alpha_F}+C\frac{(2h)^{2H}+h^{2H}}{2H}+C\frac{h^{2H}-h^2T^{2H-2}}{2-2H} \\
&\le C(h^{2\alpha_F}+h^{2H})\le Ch^{2H}. \tag*{(125)}
\end{aligned} E [ ∣ Θ t + h − Θ t ∣ 2 ] ≤ C h 2 α F + C ∫ 0 T ∣ K ( u + h ) − K ( u ) ∣ 2 d u + C ∫ 0 h ∣ K ( u ) ∣ 2 d u ≤ C h 2 α F + C ∫ 0 2 h u 2 H − 1 d u + C h 2 ∫ h T u 2 H − 3 d u + C ∫ 0 h u 2 H − 1 d u ≤ C h 2 α F + C 2 H ( 2 h ) 2 H + h 2 H + C 2 − 2 H h 2 H − h 2 T 2 H − 2 ≤ C ( h 2 α F + h 2 H ) ≤ C h 2 H . (125)
This proves (FT3). Gaussian moment equivalence gives E [ ∣ Θ t + h − Θ t ∣ p ] ≤ C p h p H \mathbb{E}[|\Theta_{t+h}-\Theta_t|^p]\le C_ph^{pH} E [ ∣ Θ t + h − Θ t ∣ p ] ≤ C p h p H for every p ≥ 2 p\ge2 p ≥ 2 . Choose p > 1 / H p>1/H p > 1/ H and apply Kolmogorov’s continuity theorem [32 Chapter 4] . Finally,
ρ ( t ) 2 / ( 1 + H ) = ∣ c K q ( t ) ∣ 2 / ( 1 + H ) Γ H 1 / ( 1 + H ) , \rho(t)^{2/(1+H)}=|c_Kq(t)|^{2/(1+H)}\Gamma_H^{1/(1+H)}, ρ ( t ) 2/ ( 1 + H ) = ∣ c K q ( t ) ∣ 2/ ( 1 + H ) Γ H 1/ ( 1 + H ) ,
so theorem 6.2 gives (64).
Bounds and positive quadrature
Single-window bounds
Proof of proposition 7.4. Lower bound. For target σ B H \sigma B^H σ B H on [ 0 , T ] [0,T] [ 0 , T ] , lemma 2.7 gives
D h : = ( T − h ) m 1 σ h H ≤ h R + 2 T M . D_h:=(T-h)m_1\sigma h^H\le hR+2\sqrt{TM}. D h := ( T − h ) m 1 σ h H ≤ h R + 2 T M .
With c : = λ b c:=\lambda b c := λb and z : = 2 T M z:=2\sqrt{TM} z := 2 T M , this bounds the objective by
J λ ( ϕ ; σ B H ) ≥ a z 2 8 T + c h ( D h − z ) + ≥ min 0 ≤ z ≤ D h { a z 2 8 T + c h ( D h − z ) } . J_\lambda(\phi;\sigma B^H)\ge\frac{az^2}{8T}+\frac{c}{h}(D_h-z)_+\ge\min_{0\le z\le D_h}\left\{\frac{az^2}{8T}+\frac{c}{h}(D_h-z)\right\}. J λ ( ϕ ; σ B H ) ≥ 8 T a z 2 + h c ( D h − z ) + ≥ 0 ≤ z ≤ D h min { 8 T a z 2 + h c ( D h − z ) } .
For z > D h z>D_h z > D h the first expression increases. On [ 0 , D h ] [0,D_h] [ 0 , D h ] , its unconstrained minimizer is 4 T c / ( a h ) 4Tc/(ah) 4 T c / ( ah ) , by differentiation. Hence
J λ ( ϕ ; σ B H ) ≥ { a D h 2 / ( 8 T ) , D h ≤ 4 T c / ( a h ) , c D h / h − 2 T c 2 / ( a h 2 ) , D h > 4 T c / ( a h ) . J_\lambda(\phi;\sigma B^H)\ge
\begin{cases}
aD_h^2/(8T), & D_h\le4Tc/(ah),\\
cD_h/h-2Tc^2/(ah^2), & D_h>4Tc/(ah).
\end{cases} J λ ( ϕ ; σ B H ) ≥ { a D h 2 / ( 8 T ) , c D h / h − 2 T c 2 / ( a h 2 ) , D h ≤ 4 T c / ( ah ) , D h > 4 T c / ( ah ) .
Let ℓ : = ( c / ( a σ ) ) 1 / ( 1 + H ) \ell:=(c/(a\sigma))^{1/(1+H)} ℓ := ( c / ( aσ ) ) 1/ ( 1 + H ) and h : = q ℓ h:=q\ell h := q ℓ , q > 0 q>0 q > 0 . The second case applies whenever
a h D h 4 T c = m 1 4 ( 1 − q ℓ T ) q H + 1 > 1. \frac{ahD_h}{4Tc}=\frac{m_1}{4}\left(1-\frac{q\ell}{T}\right)q^{H+1}>1. 4 T c ah D h = 4 m 1 ( 1 − T q ℓ ) q H + 1 > 1.
In this case, the identity c = a σ ℓ 1 + H c=a\sigma\ell^{1+H} c = aσ ℓ 1 + H gives
J λ ( ϕ ; σ B H ) T a σ 2 ℓ 2 H ≥ c D h / h − 2 T c 2 / ( a h 2 ) T a σ 2 ℓ 2 H = m 1 ( 1 − q ℓ T ) q H − 1 − 2 q − 2 \begin{aligned}
\frac{J_\lambda(\phi;\sigma B^H)}{Ta\sigma^2\ell^{2H}}\ge\frac{cD_h/h-2Tc^2/(ah^2)}{Ta\sigma^2\ell^{2H}} \\
&=m_1\left(1-\frac{q\ell}{T}\right)q^{H-1}-2q^{-2}
\end{aligned} T a σ 2 ℓ 2 H J λ ( ϕ ; σ B H ) ≥ T a σ 2 ℓ 2 H c D h / h − 2 T c 2 / ( a h 2 ) = m 1 ( 1 − T q ℓ ) q H − 1 − 2 q − 2
→ λ → 0 m 1 q H − 1 − 2 q − 2 . \xrightarrow{\lambda\to0}m_1q^{H-1}-2q^{-2}. λ → 0 m 1 q H − 1 − 2 q − 2 .
The derivative of the limiting expression is
d d q ( m 1 q H − 1 − 2 q − 2 ) = q − 3 { 4 − m 1 ( 1 − H ) q H + 1 } . \frac{\mathrm{d}}{\mathrm{d}q}\left(m_1q^{H-1}-2q^{-2}\right)=q^{-3}\{4-m_1(1-H)q^{H+1}\}. d q d ( m 1 q H − 1 − 2 q − 2 ) = q − 3 { 4 − m 1 ( 1 − H ) q H + 1 } .
It changes sign from positive to negative at
q ∗ H + 1 = 4 m 1 ( 1 − H ) , m 1 q ∗ H + 1 = 4 1 − H > 4. q_*^{H+1}=\frac{4}{m_1(1-H)},\qquad m_1q_*^{H+1}=\frac{4}{1-H}>4. q ∗ H + 1 = m 1 ( 1 − H ) 4 , m 1 q ∗ H + 1 = 1 − H 4 > 4.
Thus q ∗ q_\ast q ∗ is the global maximizer and satisfies the second-case condition for small λ \lambda λ . Using m 1 ( 1 − H ) q ∗ H + 1 = 4 m_1(1-H)q_\ast^{H+1}=4 m 1 ( 1 − H ) q ∗ H + 1 = 4 and taking infima over controls therefore gives
lim inf λ → 0 V λ ( T ; a , b , σ ) T a σ 2 ℓ 2 H ≥ m 1 q ∗ H − 1 − 2 q ∗ − 2 = m 1 ( 1 + H ) 2 q ∗ H − 1 = C l o w ( H ) . \liminf_{\lambda\to0}\frac{V_\lambda(T;a,b,\sigma)}{Ta\sigma^2\ell^{2H}} \ge m_1q_\ast^{H-1}-2q_\ast^{-2}
= \frac{m_1(1+H)}{2}q_\ast^{H-1}=C_{\mathrm{low}}(H). λ → 0 lim inf T a σ 2 ℓ 2 H V λ ( T ; a , b , σ ) ≥ m 1 q ∗ H − 1 − 2 q ∗ − 2 = 2 m 1 ( 1 + H ) q ∗ H − 1 = C low ( H ) .
By theorem 3.4, the left-hand side is κ H \kappa_H κ H .
Upper bound. On the full-past space, use the absolutely continuous rule
ψ ( h ) : { R ⟶ L 2 ( Ω ) t ⟼ 1 h ∫ t − h t B s H d s \psi^{(h)}:\begin{cases}
\mathbb{R}\longrightarrow L^2(\Omega)\\
t\longmapsto\dfrac{1}{h}\int_{t-h}^{t}\mathbf{B}_s^H\,\mathrm{d}s
\end{cases} ψ ( h ) : ⎩ ⎨ ⎧ R ⟶ L 2 ( Ω ) t ⟼ h 1 ∫ t − h t B s H d s
Its derivative and tracking error are
ψ ˙ t ( h ) = B t H − B t − h H h , \dot{\psi}_t^{(h)}=\frac{\mathbf{B}_t^H-\mathbf{B}_{t-h}^H}{h}, ψ ˙ t ( h ) = h B t H − B t − h H ,
and
B t H − ψ t ( h ) = 1 h ∫ 0 h ( B t H − B t − u H ) d u . \mathbf{B}_t^H-\psi_t^{(h)}=\frac{1}{h}\int_0^h\left(\mathbf{B}_t^H-\mathbf{B}_{t-u}^H\right)\,\mathrm{d}u. B t H − ψ t ( h ) = h 1 ∫ 0 h ( B t H − B t − u H ) d u .
Stationary increments and Fubini give
∫ 0 h ∫ 0 h ∣ u − v ∣ 2 H d u d v = 2 ∫ 0 h ∫ 0 u ( u − v ) 2 H d v d u = 2 h 2 H + 2 ( 2 H + 1 ) ( 2 H + 2 ) . \int_0^h\int_0^h |u-v|^{2H}\,\mathrm{d}u\,\mathrm{d}v
=2\int_0^h\int_0^u(u-v)^{2H}\,\mathrm{d}v\,\mathrm{d}u
=\frac{2h^{2H+2}}{(2H+1)(2H+2)}. ∫ 0 h ∫ 0 h ∣ u − v ∣ 2 H d u d v = 2 ∫ 0 h ∫ 0 u ( u − v ) 2 H d v d u = ( 2 H + 1 ) ( 2 H + 2 ) 2 h 2 H + 2 .
Then,
E [ ∣ B t H − ψ t ( h ) ∣ 2 ] = 1 h 2 ∫ 0 h ∫ 0 h R H ( u , v ) d u d v = h 2 H 2 H + 1 − h 2 H ( 2 H + 1 ) ( 2 H + 2 ) = h 2 H 2 ( H + 1 ) , \mathbb{E}\left[\left|\mathbf{B}_t^H-\psi_t^{(h)}\right|^2\right]
=\frac{1}{h^2}\int_0^h\int_0^h R_H(u,v)\,\mathrm{d}u\,\mathrm{d}v
=\frac{h^{2H}}{2H+1}-\frac{h^{2H}}{(2H+1)(2H+2)}
=\frac{h^{2H}}{2(H+1)}, E [ B t H − ψ t ( h ) 2 ] = h 2 1 ∫ 0 h ∫ 0 h R H ( u , v ) d u d v = 2 H + 1 h 2 H − ( 2 H + 1 ) ( 2 H + 2 ) h 2 H = 2 ( H + 1 ) h 2 H ,
and
E [ ∣ ψ ˙ t ( h ) ∣ ] = m 1 h H − 1 . \mathbb{E}\left[\left|\dot{\psi}_t^{(h)}\right|\right]=m_1h^{H-1}. E [ ψ ˙ t ( h ) ] = m 1 h H − 1 .
The initial atom and terminal reset each have expectation m 1 h H / 2 ( H + 1 ) m_1h^H/\sqrt{2(H+1)} m 1 h H / 2 ( H + 1 ) , so their contribution per unit time vanishes. The long-run cost is
C : { ( 0 , ∞ ) ⟶ ( 0 , ∞ ) h ⟼ h 2 H 4 ( H + 1 ) + m 1 h H − 1 C:\begin{cases}
(0,\infty)\longrightarrow(0,\infty)\\
h\longmapsto\dfrac{h^{2H}}{4(H+1)}+m_1h^{H-1}
\end{cases} C : ⎩ ⎨ ⎧ ( 0 , ∞ ) ⟶ ( 0 , ∞ ) h ⟼ 4 ( H + 1 ) h 2 H + m 1 h H − 1
with derivative
C ′ ( h ) = h H − 2 { H 2 ( H + 1 ) h H + 1 − m 1 ( 1 − H ) } . C'(h)=h^{H-2}\left\{\frac{H}{2(H+1)}h^{H+1}-m_1(1-H)\right\}. C ′ ( h ) = h H − 2 { 2 ( H + 1 ) H h H + 1 − m 1 ( 1 − H ) } .
The derivative changes sign once, at
h ∗ H + 1 = 2 m 1 ( 1 − H ) ( 1 + H ) H , m 1 h ∗ H − 1 = H h ∗ 2 H 2 ( 1 − H ) ( 1 + H ) . h_\ast^{H+1}=\frac{2m_1(1-H)(1+H)}{H},\qquad
m_1h_\ast^{H-1}=\frac{Hh_\ast^{2H}}{2(1-H)(1+H)}. h ∗ H + 1 = H 2 m 1 ( 1 − H ) ( 1 + H ) , m 1 h ∗ H − 1 = 2 ( 1 − H ) ( 1 + H ) H h ∗ 2 H .
Substitution gives
min h > 0 C ( h ) = h ∗ 2 H 4 ( 1 − H ) = 1 4 ( 1 − H ) ( 2 m 1 ( 1 − H ) ( 1 + H ) H ) 2 H / ( 1 + H ) = C m a ( H ) . \min_{h>0}C(h)=\frac{h_\ast^{2H}}{4(1-H)}
=\frac{1}{4(1-H)}
\left(\frac{2m_1(1-H)(1+H)}{H}\right)^{2H/(1+H)}
=C_{\mathrm{ma}}(H). h > 0 min C ( h ) = 4 ( 1 − H ) h ∗ 2 H = 4 ( 1 − H ) 1 ( H 2 m 1 ( 1 − H ) ( 1 + H ) ) 2 H / ( 1 + H ) = C ma ( H ) .
Apply theorem 5.3 to conclude.
Multi-window bounds
Fix an integer m ≥ 1 m \ge1 m ≥ 1 and define the affine weight space
W m : = { w ∈ R m : ∑ i = 1 m w i = 1 } . \mathscr{W}_m := \left\{ w \in\mathbb{R}^m : \sum_{i=1}^{m} w_i = 1 \right\}. W m := { w ∈ R m : i = 1 ∑ m w i = 1 } .
For q = ( q 1 , … , q m ) ∈ ( 0 , ∞ ) m q = (q_1,\ldots,q_m) \in(0,\infty)^m q = ( q 1 , … , q m ) ∈ ( 0 , ∞ ) m and w ∈ W m w \in\mathscr{W}_m w ∈ W m , define on the two-sided fBm space
ψ w , q : ∣ R ⟶ L 2 ( Ω ) t ⟼ ∑ i = 1 m w i q i ∫ t − q i t B s H d s (126) \psi^{w,q}:
\left|
\begin{aligned}
\mathbb{R} &\longrightarrow L^2(\Omega)\\
t &\longmapsto\sum_{i=1}^{m} \frac{w_i}{q_i}\int_{t-q_i}^{t} B_s^H\,\mathrm{d}s
\end{aligned}
\right. \tag*{(126)} ψ w , q : R t ⟶ L 2 ( Ω ) ⟼ i = 1 ∑ m q i w i ∫ t − q i t B s H d s (126)
Define the symmetric matrices A H ( q ) , D H ( q ) ∈ R m × m A^H(q),D^H(q) \in\mathbb{R}^{m \times m} A H ( q ) , D H ( q ) ∈ R m × m by
A H : ∣ ( 0 , ∞ ) m ⟶ R m × m q ⟼ [ 1 q i q j ∫ 0 q i ∫ 0 q j R H ( u , v ) d u d v ] i , j = 1 m (127) A^H:
\left|
\begin{aligned}
(0,\infty)^m &\longrightarrow\mathbb{R}^{m \times m}\\
q &\longmapsto\left[\frac{1}{q_iq_j}\int_0^{q_i}\int_0^{q_j} R_H(u,v)\,\mathrm{d}u\,\mathrm{d}v\right]_{i,j=1}^{m}
\end{aligned}
\right. \tag*{(127)} A H : ( 0 , ∞ ) m q ⟶ R m × m ⟼ [ q i q j 1 ∫ 0 q i ∫ 0 q j R H ( u , v ) d u d v ] i , j = 1 m (127)
and
D H : ∣ ( 0 , ∞ ) m ⟶ R m × m q ⟼ [ R H ( q i , q j ) q i q j ] i , j = 1 m (128) D^H:
\left|
\begin{aligned}
(0,\infty)^m &\longrightarrow\mathbb{R}^{m \times m}\\
q &\longmapsto\left[\frac{R_H(q_i,q_j)}{q_iq_j}\right]_{i,j=1}^{m}
\end{aligned}
\right. \tag*{(128)} D H : ( 0 , ∞ ) m q ⟶ R m × m ⟼ [ q i q j R H ( q i , q j ) ] i , j = 1 m (128)
Both matrices are positive semidefinite because they are covariance matrices of Gaussian linear functionals. Their quadratic forms are
A H : ∣ W m × ( 0 , ∞ ) m ⟶ [ 0 , ∞ ) ( w , q ) ⟼ w T A H ( q ) w A_H:
\left|
\begin{aligned}
\mathscr{W}_m \times(0,\infty)^m &\longrightarrow[0,\infty)\\
(w,q) &\longmapsto w^{\mathsf{T}}A^H(q)w
\end{aligned}
\right. A H : W m × ( 0 , ∞ ) m ( w , q ) ⟶ [ 0 , ∞ ) ⟼ w T A H ( q ) w
and
D H : ∣ W m × ( 0 , ∞ ) m ⟶ [ 0 , ∞ ) ( w , q ) ⟼ w T D H ( q ) w D_H:
\left|
\begin{aligned}
\mathscr{W}_m \times(0,\infty)^m &\longrightarrow[0,\infty)\\
(w,q) &\longmapsto w^{\mathsf{T}}D^H(q)w
\end{aligned}
\right. D H : W m × ( 0 , ∞ ) m ( w , q ) ⟶ [ 0 , ∞ ) ⟼ w T D H ( q ) w
For a , b > 0 a,b>0 a , b > 0 , integrating the covariance gives
∫ 0 a ∫ 0 b R H ( u , v ) d u d v = 1 2 ∫ 0 a ∫ 0 b ( u 2 H + v 2 H − ∣ u − v ∣ 2 H ) d v d u = b a 2 H + 1 + a b 2 H + 1 2 ( 2 H + 1 ) − a 2 H + 2 + b 2 H + 2 − ∣ a − b ∣ 2 H + 2 2 ( 2 H + 1 ) ( 2 H + 2 ) . (129) \begin{aligned}
\int_0^a\int_0^b R_H(u,v)\,\mathrm{d}u\,\mathrm{d}v
&= \frac{1}{2}\int_0^a\int_0^b \left(u^{2H}+v^{2H}-|u-v|^{2H}\right)\,\mathrm{d}v\,\mathrm{d}u\\
&= \frac{ba^{2H+1}+ab^{2H+1}}{2(2H+1)}
-\frac{a^{2H+2}+b^{2H+2}-|a-b|^{2H+2}}{2(2H+1)(2H+2)}. \tag*{(129)}
\end{aligned} ∫ 0 a ∫ 0 b R H ( u , v ) d u d v = 2 1 ∫ 0 a ∫ 0 b ( u 2 H + v 2 H − ∣ u − v ∣ 2 H ) d v d u = 2 ( 2 H + 1 ) b a 2 H + 1 + a b 2 H + 1 − 2 ( 2 H + 1 ) ( 2 H + 2 ) a 2 H + 2 + b 2 H + 2 − ∣ a − b ∣ 2 H + 2 . (129)
Proposition I.1 (multi-window upper bound). Define
C H : ∣ W m × ( 0 , ∞ ) m ⟶ [ 0 , ∞ ) ( w , q ) ⟼ 1 2 A H ( w , q ) + m 1 D H ( w , q ) C_H:
\left|
\begin{aligned}
\mathscr{W}_m \times(0,\infty)^m &\longrightarrow[0,\infty)\\
(w,q) &\longmapsto\frac{1}{2}A_H(w,q)+m_1\sqrt{D_H(w,q)}
\end{aligned}
\right. C H : W m × ( 0 , ∞ ) m ( w , q ) ⟶ [ 0 , ∞ ) ⟼ 2 1 A H ( w , q ) + m 1 D H ( w , q ) For every q ∈ ( 0 , ∞ ) m q \in(0,\infty)^m q ∈ ( 0 , ∞ ) m and every w ∈ W m w \in\mathscr{W}_m w ∈ W m ,
κ H ≤ C H ( w , q ) . (130) \kappa_H \le C_H(w,q). \tag*{(130)} κ H ≤ C H ( w , q ) . (130)
Proof. Since ∑ i = 1 m w i = 1 \sum_{i=1}^{m} w_i=1 ∑ i = 1 m w i = 1 ,
B t H − ψ t w , q = ∑ i = 1 m w i q i ∫ 0 q i ( B t H − B t − u H ) d u , B_t^H-\psi_t^{w,q}
=\sum_{i=1}^{m}\frac{w_i}{q_i}\int_0^{q_i}\left(B_t^H-B_{t-u}^H\right)\,\mathrm{d}u, B t H − ψ t w , q = i = 1 ∑ m q i w i ∫ 0 q i ( B t H − B t − u H ) d u ,
and
ψ ˙ t w , q = ∑ i = 1 m w i q i ( B t H − B t − q i H ) . \dot{\psi}_t^{w,q}=\sum_{i=1}^{m}\frac{w_i}{q_i}\left(B_t^H-B_{t-q_i}^H\right). ψ ˙ t w , q = i = 1 ∑ m q i w i ( B t H − B t − q i H ) .
These variables are centered Gaussian, and stationarity of the increments gives their variances:
E [ ∣ B t H − ψ t w , q ∣ 2 ] = ∑ i = 1 m ∑ j = 1 m w i w j q i q j ∫ 0 q i ∫ 0 q j R H ( u , v ) d v d u = w T A H ( q ) w = A H ( w , q ) . \begin{aligned}
\mathbb{E}\left[\left|B_t^H-\psi_t^{w,q}\right|^2\right]
&=\sum_{i=1}^{m}\sum_{j=1}^{m}\frac{w_iw_j}{q_iq_j}\int_0^{q_i}\int_0^{q_j}R_H(u,v)\,\mathrm{d}v\,\mathrm{d}u\\
&=w^{\mathsf{T}}A^H(q)w=A_H(w,q).
\end{aligned} E [ B t H − ψ t w , q 2 ] = i = 1 ∑ m j = 1 ∑ m q i q j w i w j ∫ 0 q i ∫ 0 q j R H ( u , v ) d v d u = w T A H ( q ) w = A H ( w , q ) .
and
E [ ∣ ψ t w , q ∣ 2 ] = ∑ i = 1 m ∑ j = 1 m w i w j q i q j R H ( q i , q j ) = w ⊤ D H ( q ) w = D H ( w , q ) . \mathbb{E}\left[\left|\psi_t^{w,q}\right|^2\right]=\sum_{i=1}^{m}\sum_{j=1}^{m}\frac{w_iw_j}{q_iq_j}R_H(q_i,q_j)=w^\top D^H(q)w=D_H(w,q). E [ ∣ ψ t w , q ∣ 2 ] = i = 1 ∑ m j = 1 ∑ m q i q j w i w j R H ( q i , q j ) = w ⊤ D H ( q ) w = D H ( w , q ) .
The initial atom and terminal reset each have expectation m 1 A H ( w , q ) m_1\sqrt{A_H(w,q)} m 1 A H ( w , q ) . Therefore the full-past reset cost is
J L B H ( ψ w , q ) = L ( 1 2 A H ( w , q ) + m 1 D H ( w , q ) ) + 2 m 1 A H ( w , q ) . J_L^{\mathrm{BH}}(\psi^{w,q})=L\left(\frac{1}{2}A_H(w,q)+m_1\sqrt{D_H(w,q)}\right)+2m_1\sqrt{A_H(w,q)}. J L BH ( ψ w , q ) = L ( 2 1 A H ( w , q ) + m 1 D H ( w , q ) ) + 2 m 1 A H ( w , q ) .
Divide by L L L and use u H p a s t ( L ) ≤ J L B H ( ψ w , q ) u_H^{\mathrm{past}}(L)\leq J_L^{\mathrm{BH}}(\psi^{w,q}) u H past ( L ) ≤ J L BH ( ψ w , q ) and theorem 5.3.
Corollary I.2 (scale-free multi-window bound). Fix m ≥ 1 m\geq1 m ≥ 1 , w ∈ W m w\in\mathcal{W}_m w ∈ W m , and q ∈ ( 0 , ∞ ) m q\in(0,\infty)^m q ∈ ( 0 , ∞ ) m . Assume A H ( w , q ) > 0 A_H(w,q)>0 A H ( w , q ) > 0 and D H ( w , q ) > 0 D_H(w,q)>0 D H ( w , q ) > 0 . For c > 0 c>0 c > 0 ,
A H ( w , c q ) = c 2 H A H ( w , q ) , D H ( w , c q ) = c 2 H − 2 D H ( w , q ) . (131) A_H(w,cq)=c^{2H}A_H(w,q),\qquad D_H(w,cq)=c^{2H-2}D_H(w,q). \tag*{(131)} A H ( w , c q ) = c 2 H A H ( w , q ) , D H ( w , c q ) = c 2 H − 2 D H ( w , q ) . (131) The unique minimizer of c ↦ C H ( w , c q ) c\mapsto C_H(w,cq) c ↦ C H ( w , c q ) is
c H ∗ ( w , q ) : = ( m 1 ( 1 − H ) D H ( w , q ) H A H ( w , q ) ) 1 / ( 1 + H ) . (132) c_H^*(w,q):=\left(\frac{m_1(1-H)\sqrt{D_H(w,q)}}{HA_H(w,q)}\right)^{1/(1+H)}. \tag*{(132)} c H ∗ ( w , q ) := ( H A H ( w , q ) m 1 ( 1 − H ) D H ( w , q ) ) 1/ ( 1 + H ) . (132) Hence
κ H ≤ C ‾ H ( w , q ) , (133) \kappa_H\leq\overline{C}_H(w,q), \tag*{(133)} κ H ≤ C H ( w , q ) , (133) where
C ‾ H ( w , q ) : = 1 + H 2 ( 1 − H ) A H ( w , q ) ( 1 − H ) / ( 1 + H ) ( m 1 ( 1 − H ) H D H ( w , q ) ) 2 H / ( 1 + H ) . (134) \overline{C}_H(w,q):=\frac{1+H}{2(1-H)}A_H(w,q)^{(1-H)/(1+H)}
\left(\frac{m_1(1-H)}{H}\sqrt{D_H(w,q)}\right)^{2H/(1+H)}. \tag*{(134)} C H ( w , q ) := 2 ( 1 − H ) 1 + H A H ( w , q ) ( 1 − H ) / ( 1 + H ) ( H m 1 ( 1 − H ) D H ( w , q ) ) 2 H / ( 1 + H ) . (134) The quantity C ‾ H ( w , q ) \overline{C}_H(w,q) C H ( w , q ) does not change when q q q is replaced by c q cq c q . For m = 1 m=1 m = 1 , w 1 = q 1 = 1 w_1=q_1=1 w 1 = q 1 = 1 , it is equal to C m a ( H ) C_{\mathrm{ma}}(H) C ma ( H ) .
Proof. From R H ( c u , c v ) = c 2 H R H ( u , v ) R_H(cu,cv)=c^{2H}R_H(u,v) R H ( c u , c v ) = c 2 H R H ( u , v ) and the substitutions u = c u ~ u=c\tilde{u} u = c u ~ , v = c v ~ v=c\tilde{v} v = c v ~ ,
A i j H ( c q ) = c 2 H + 2 c 2 q i q j ∫ 0 q i ∫ 0 q j R H ( u ~ , v ~ ) d u ~ d v ~ = c 2 H A i j H ( q ) , A_{ij}^H(cq)=\frac{c^{2H+2}}{c^2q_iq_j}\int_0^{q_i}\int_0^{q_j}R_H(\tilde{u},\tilde{v})\,\mathrm{d}\tilde{u}\,\mathrm{d}\tilde{v}=c^{2H}A_{ij}^H(q), A ij H ( c q ) = c 2 q i q j c 2 H + 2 ∫ 0 q i ∫ 0 q j R H ( u ~ , v ~ ) d u ~ d v ~ = c 2 H A ij H ( q ) ,
and
D i j H ( c q ) = c 2 H − 2 D i j H ( q ) . D_{ij}^H(cq)=c^{2H-2}D_{ij}^H(q). D ij H ( c q ) = c 2 H − 2 D ij H ( q ) .
Write A : = A H ( w , q ) > 0 A:=A_H(w,q)>0 A := A H ( w , q ) > 0 and D : = D H ( w , q ) > 0 D:=D_H(w,q)>0 D := D H ( w , q ) > 0 . The scaling identities then give
C H ( w , c q ) = 1 2 A c 2 H + m 1 D c H − 1 . C_H(w,cq)=\frac{1}{2}Ac^{2H}+m_1\sqrt{D}\,c^{H-1}. C H ( w , c q ) = 2 1 A c 2 H + m 1 D c H − 1 .
Then,
d d c C H ( w , c q ) = c H − 2 { H A c H + 1 − m 1 ( 1 − H ) D } . \frac{\mathrm{d}}{\mathrm{d}c}C_H(w,cq)=c^{H-2}\left\{HAc^{H+1}-m_1(1-H)\sqrt{D}\right\}. d c d C H ( w , c q ) = c H − 2 { H A c H + 1 − m 1 ( 1 − H ) D } .
The bracket is strictly increasing from negative to positive. Its zero is (132) . At c = c H ∗ ( w , q ) c=c_H^*(w,q) c = c H ∗ ( w , q ) ,
m 1 D c H − 1 = H 1 − H A c 2 H , m_1\sqrt{D}\,c^{H-1}=\frac{H}{1-H}Ac^{2H}, m 1 D c H − 1 = 1 − H H A c 2 H ,
so
C H ( w , c q ) = 1 + H 2 ( 1 − H ) A c 2 H = 1 + H 2 ( 1 − H ) A ( 1 − H ) / ( 1 + H ) ( m 1 ( 1 − H ) H D ) 2 H / ( 1 + H ) . C_H(w,cq)=\frac{1+H}{2(1-H)}Ac^{2H}
=\frac{1+H}{2(1-H)}A^{(1-H)/(1+H)}
\left(\frac{m_1(1-H)}{H}\sqrt{D}\right)^{2H/(1+H)}. C H ( w , c q ) = 2 ( 1 − H ) 1 + H A c 2 H = 2 ( 1 − H ) 1 + H A ( 1 − H ) / ( 1 + H ) ( H m 1 ( 1 − H ) D ) 2 H / ( 1 + H ) .
Apply proposition I.1. Under q ↦ c q q\mapsto cq q ↦ c q , the total power of c c c in this expression is
2 H ( 1 − H ) 1 + H + 2 H ( H − 1 ) 1 + H = 0 , \frac{2H(1-H)}{1+H}+\frac{2H(H-1)}{1+H}=0, 1 + H 2 H ( 1 − H ) + 1 + H 2 H ( H − 1 ) = 0 ,
which proves scale invariance. For m = 1 m=1 m = 1 and w 1 = q 1 = 1 w_1=q_1=1 w 1 = q 1 = 1 , A = 1 / [ 2 ( H + 1 ) ] A=1/[2(H+1)] A = 1/ [ 2 ( H + 1 )] and D = 1 D=1 D = 1 . Substitution gives C m a ( H ) C_{\mathrm{ma}}(H) C ma ( H ) .
Proposition I.3 (optimal weights for fixed windows). Fix q q q and assume D = D H ( q ) D = D^{H}(q) D = D H ( q ) is positive definite. Put A = A H ( q ) A = A^{H}(q) A = A H ( q ) , 1 = ( 1 , … , 1 ) T \mathbf{1} = (1,\ldots,1)^{\mathsf{T}} 1 = ( 1 , … , 1 ) T , and, for r > 0 r > 0 r > 0 ,
M : { ( 0 , ∞ ) ⟶ R m × m r ⟼ A + m 1 r D M:\left\{\begin{array}{rcl}
(0,\infty)&\longrightarrow&\mathbb{R}^{m\times m}\\
r&\longmapsto&A+\dfrac{m_1}{r}D
\end{array}\right. M : { ( 0 , ∞ ) r ⟶ ⟼ R m × m A + r m 1 D Then the best bound within these windows is
min w ∈ W m C H ( w , q ) = min r > 0 { 1 1 T M ( r ) − 1 1 + m 1 r 2 } . (135) \min_{w\in\mathcal{W}_m} C_H(w,q)=\min_{r>0}\left\{\frac{1}{\mathbf{1}^{\mathsf{T}}M(r)^{-1}\mathbf{1}}+\frac{m_1r}{2}\right\}. \tag*{(135)} w ∈ W m min C H ( w , q ) = r > 0 min { 1 T M ( r ) − 1 1 1 + 2 m 1 r } . (135) The function minimized on the right-hand side is convex in r r r . If r ∗ r_* r ∗ is a minimizer, the optimal weight vector is
w ∗ : = M ( r ∗ ) − 1 1 1 T M ( r ∗ ) − 1 1 , r ∗ 2 = w ∗ T D w ∗ . (136) w_*:=\frac{M(r_*)^{-1}\mathbf{1}}{\mathbf{1}^{\mathsf{T}}M(r_*)^{-1}\mathbf{1}},\qquad r_*^2=w_*^{\mathsf{T}}Dw_*. \tag*{(136)} w ∗ := 1 T M ( r ∗ ) − 1 1 M ( r ∗ ) − 1 1 , r ∗ 2 = w ∗ T D w ∗ . (136)
Proof. For z > 0 z>0 z > 0 and r > 0 r>0 r > 0 , completing the square gives
z 2 r + r 2 − z = ( r − z ) 2 2 r ≥ 0 , \frac{z}{2r}+\frac{r}{2}-\sqrt{z}=\frac{(r-\sqrt{z})^2}{2r}\geq0, 2 r z + 2 r − z = 2 r ( r − z ) 2 ≥ 0 ,
with equality exactly when r 2 = z r^2=z r 2 = z . Define
F : { W m × ( 0 , ∞ ) ⟶ [ 0 , ∞ ) ( w , r ) ⟼ 1 2 w T M ( r ) w + 1 2 m 1 r F:\left\{\begin{array}{rcl}
\mathcal{W}_m\times(0,\infty)&\longrightarrow&[0,\infty)\\
(w,r)&\longmapsto&\frac{1}{2}w^{\mathsf{T}}M(r)w+\frac{1}{2}m_1r
\end{array}\right. F : { W m × ( 0 , ∞ ) ( w , r ) ⟶ ⟼ [ 0 , ∞ ) 2 1 w T M ( r ) w + 2 1 m 1 r
This identity gives C H ( w , q ) = min r > 0 F ( w , r ) C_H(w,q)=\min_{r>0}F(w,r) C H ( w , q ) = min r > 0 F ( w , r ) . For fixed r r r , M ( r ) M(r) M ( r ) is positive definite. A Lagrange multiplier ζ ∈ R \zeta\in\mathbb{R} ζ ∈ R for 1 T w = 1 \mathbf{1}^{\mathsf{T}}w=1 1 T w = 1 gives
M ( r ) w = ζ 1 , w ( r ) = M ( r ) − 1 1 1 T M ( r ) − 1 1 . M(r)w=\zeta\mathbf{1},\qquad w(r)=\frac{M(r)^{-1}\mathbf{1}}{\mathbf{1}^{\mathsf{T}}M(r)^{-1}\mathbf{1}}. M ( r ) w = ζ 1 , w ( r ) = 1 T M ( r ) − 1 1 M ( r ) − 1 1 .
Thus
1 2 w ( r ) T M ( r ) w ( r ) = ζ 2 1 T w ( r ) = 1 21 T M ( r ) − 1 1 , \frac{1}{2}w(r)^{\mathsf{T}}M(r)w(r)=\frac{\zeta}{2}\mathbf{1}^{\mathsf{T}}w(r)=\frac{1}{2\mathbf{1}^{\mathsf{T}}M(r)^{-1}\mathbf{1}}, 2 1 w ( r ) T M ( r ) w ( r ) = 2 ζ 1 T w ( r ) = 2 1 T M ( r ) − 1 1 1 ,
which proves (135) after minimizing over r r r . For directions z ∈ R m z\in\mathbb{R}^m z ∈ R m and s ∈ R s\in\mathbb{R} s ∈ R , the second directional variation of the quadratic perspective is
d 2 ( w T D w 2 r ) [ ( z , s ) , ( z , s ) ] = 1 r ( z − s r w ) T D ( z − s r w ) ≥ 0. \mathrm{d}^{2}\left(\frac{w^{\mathsf{T}}Dw}{2r}\right)[(z,s),(z,s)]
=\frac{1}{r}\left(z-\frac{s}{r}w\right)^{\mathsf{T}}D\left(z-\frac{s}{r}w\right)\geq0. d 2 ( 2 r w T D w ) [( z , s ) , ( z , s )] = r 1 ( z − r s w ) T D ( z − r s w ) ≥ 0.
Hence F F F is jointly convex, and partial minimization on 1 T w = 1 \mathbf{1}^{\mathsf{T}}w=1 1 T w = 1 preserves convexity. Since D D D is positive definite and ∥ w ∥ 2 ≥ 1 / m \lVert w\rVert^{2}\geq1/m ∥ w ∥ 2 ≥ 1/ m on W m \mathcal{W}_m W m ,
F ( w , r ) ≥ m 1 2 r w T D w + m 1 r 2 ≥ m 1 λ min ( D ) 2 m r + m 1 r 2 → r → 0 o r r → ∞ ∞ , F(w,r)\geq\frac{m_1}{2r}w^{\mathsf{T}}Dw+\frac{m_1r}{2}
\geq\frac{m_1\lambda_{\min}(D)}{2mr}+\frac{m_1r}{2}\xrightarrow[r\to0\ \mathrm{or}\ r\to\infty]{}\infty, F ( w , r ) ≥ 2 r m 1 w T D w + 2 m 1 r ≥ 2 m r m 1 λ m i n ( D ) + 2 m 1 r r → 0 or r → ∞ ∞ ,
where λ min ( D ) > 0 \lambda_{\min}(D)>0 λ m i n ( D ) > 0 is the smallest eigenvalue. A minimizer r ∗ r_* r ∗ exists. Equality in the scalar identity gives r ∗ 2 = w ∗ T D w ∗ r_*^2=w_*^{\mathsf{T}}Dw_* r ∗ 2 = w ∗ T D w ∗ , with w ∗ = w ( r ∗ ) w_*=w(r_*) w ∗ = w ( r ∗ ) .
Positive quadrature
Proof of theorem 8.4. OU increment kernels. For x > 0 x>0 x > 0 and t ≥ 0 t\geq0 t ≥ 0 , define
g t , x : { R ⟶ R s ⟼ e − x ( t − s ) 1 { s < t } − e x s 1 { s < 0 } g_{t,x}:\left\{\begin{array}{rcl}
\mathbb{R}&\longrightarrow&\mathbb{R}\\
s&\longmapsto&e^{-x(t-s)}\mathbf{1}_{\{s<t\}}-e^{xs}\mathbf{1}_{\{s<0\}}
\end{array}\right. g t , x : { R s ⟶ ⟼ R e − x ( t − s ) 1 { s < t } − e x s 1 { s < 0 }
For x , y > 0 x,y>0 x , y > 0 , splitting at s = 0 s=0 s = 0 gives
⟨ g t , x , g t , y ⟩ = ∫ 0 t e − ( x + y ) r d r + ( e − x t − 1 ) ( e − y t − 1 ) ∫ 0 ∞ e − ( x + y ) r d r = 1 − e − ( x + y ) t + ( e − x t − 1 ) ( e − y t − 1 ) x + y = 2 − e − x t − e − y t x + y . \begin{aligned}
\langle g_{t,x},g_{t,y}\rangle
&=\int_{0}^{t}e^{-(x+y)r}\,\mathrm{d}r+(e^{-xt}-1)(e^{-yt}-1)\int_{0}^{\infty}e^{-(x+y)r}\,\mathrm{d}r\\
&=\frac{1-e^{-(x+y)t}+(e^{-xt}-1)(e^{-yt}-1)}{x+y}\\
&=\frac{2-e^{-xt}-e^{-yt}}{x+y}.
\end{aligned} ⟨ g t , x , g t , y ⟩ = ∫ 0 t e − ( x + y ) r d r + ( e − x t − 1 ) ( e − y t − 1 ) ∫ 0 ∞ e − ( x + y ) r d r = x + y 1 − e − ( x + y ) t + ( e − x t − 1 ) ( e − y t − 1 ) = x + y 2 − e − x t − e − y t .
Setting y = x y=x y = x and using 1 − e − v ≤ v ∧ 1 1-e^{-v}\le v\wedge1 1 − e − v ≤ v ∧ 1 for v ≥ 0 v\ge0 v ≥ 0 yields
∥ g t , x ∥ L 2 ( R ) 2 = 1 − e − x t x ≤ t ∧ x − 1 . (137) \lVert g_{t,x}\rVert_{L^2(\mathbb{R})}^2=\frac{1-e^{-xt}}{x}\le t\wedge x^{-1}. \tag*{(137)} ∥ g t , x ∥ L 2 ( R ) 2 = x 1 − e − x t ≤ t ∧ x − 1 . (137)
The map x ↦ g t , x x\mapsto g_{t,x} x ↦ g t , x is differentiable in L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) . Differentiating the preceding inner product in both variables gives
∥ ∂ x g t , x ∥ 2 2 = ∂ x ∂ y 2 − e − x t − e − y t x + y ∣ y = x = − t e − x t 2 x 2 + 1 − e − x t 2 x 3 . \begin{aligned}
\lVert\partial_x g_{t,x}\rVert_2^2=\left.\partial_x\partial_y\frac{2-e^{-xt}-e^{-yt}}{x+y}\right|_{y=x} \\
&=-\frac{te^{-xt}}{2x^2}+\frac{1-e^{-xt}}{2x^3}.
\end{aligned} ∥ ∂ x g t , x ∥ 2 2 = ∂ x ∂ y x + y 2 − e − x t − e − y t y = x = − 2 x 2 t e − x t + 2 x 3 1 − e − x t .
Consequently,
∥ ∂ x g t , x ∥ L 2 ( R ) 2 = 1 − ( 1 + x t ) e − x t 2 x 3 , (138) \lVert\partial_x g_{t,x}\rVert_{L^2(\mathbb{R})}^2=\frac{1-(1+xt)e^{-xt}}{2x^3}, \tag*{(138)} ∥ ∂ x g t , x ∥ L 2 ( R ) 2 = 2 x 3 1 − ( 1 + x t ) e − x t , (138)
and
x ∥ ∂ x g t , x ∥ 2 ≤ 2 − 1 / 2 ∥ g t , x ∥ 2 , (139) x\lVert\partial_x g_{t,x}\rVert_2\le2^{-1/2}\lVert g_{t,x}\rVert_2, \tag*{(139)} x ∥ ∂ x g t , x ∥ 2 ≤ 2 − 1/2 ∥ g t , x ∥ 2 , (139)
because $0\le 1-(1+v)e^{-v}\le 1-e^{-v}f o r for f or v\ge 0.
Bochner representation and truncation. Let μ H ( d x ) : = 2 H x β − 1 d x / Γ ( β ) \mu_H(\mathrm{d}x):=\sqrt{2H}x^{\beta-1}\mathrm{d}x/\Gamma(\beta) μ H ( d x ) := 2 H x β − 1 d x /Γ ( β ) . For t > 0 t>0 t > 0 , splitting at x = t − 1 x=t^{-1} x = t − 1 gives
∫ 0 ∞ ∥ g t , x ∥ 2 μ H ( d x ) ≤ C H { t ∫ 0 1 / t x β − 1 d x + ∫ 1 / t ∞ x β − 3 / 2 d x } = C H ( 1 β + 1 H ) t H < ∞ . \begin{aligned}
\int_0^\infty\lVert g_{t,x}\rVert_2\mu_H(\mathrm{d}x)\le C_H\left\{\sqrt{t}\int_0^{1/t}x^{\beta-1}\mathrm{d}x+\int_{1/t}^\infty x^{\beta-3/2}\mathrm{d}x\right\} \\
&=C_H\left(\frac{1}{\beta}+\frac{1}{H}\right)t^H<\infty.
\end{aligned} ∫ 0 ∞ ∥ g t , x ∥ 2 μ H ( d x ) ≤ C H { t ∫ 0 1/ t x β − 1 d x + ∫ 1/ t ∞ x β − 3/2 d x } = C H ( β 1 + H 1 ) t H < ∞.
Hence the Bochner integral in L 2 ( R ) L^2(\mathbb{R}) L 2 ( R ) exists, and stochastic Fubini identifies the fBm increment kernel as
1 A H ∫ 0 ∞ g t , x μ H ( d x ) . \frac{1}{A_H}\int_0^\infty g_{t,x}\mu_H(\mathrm{d}x). A H 1 ∫ 0 ∞ g t , x μ H ( d x ) .
For t ≤ L t\le L t ≤ L , the low- and high-frequency errors satisfy
∥ ∫ 0 ε g t , x μ H ( d x ) ∥ 2 ≤ C H L ∫ 0 ε x β − 1 d x = C H β L ε β , \left\lVert\int_0^\varepsilon g_{t,x}\mu_H(\mathrm{d}x)\right\rVert_2\le C_H\sqrt{L}\int_0^\varepsilon x^{\beta-1}\mathrm{d}x=\frac{C_H}{\beta}\sqrt{L}\varepsilon^\beta, ∫ 0 ε g t , x μ H ( d x ) 2 ≤ C H L ∫ 0 ε x β − 1 d x = β C H L ε β ,
and
∥ ∫ R ∞ g t , x μ H ( d x ) ∥ 2 ≤ C H ∫ R ∞ x β − 3 / 2 d x = C H H R − H . \left\lVert\int_R^\infty g_{t,x}\mu_H(\mathrm{d}x)\right\rVert_2\le C_H\int_R^\infty x^{\beta-3/2}\mathrm{d}x=\frac{C_H}{H}R^{-H}. ∫ R ∞ g t , x μ H ( d x ) 2 ≤ C H ∫ R ∞ x β − 3/2 d x = H C H R − H .
Positive quadrature error. For x ∈ [ x j , x j + 1 ] x\in[x_j,x_{j+1}] x ∈ [ x j , x j + 1 ] , apply the fundamental theorem of calculus and the derivative estimate. Since x / x j ≤ 1 + η ≤ 2 x/x_j\le1+\eta\le2 x / x j ≤ 1 + η ≤ 2 and log ( x / x j ) ≤ η \log(x/x_j)\le\eta log ( x / x j ) ≤ η , this gives
∥ g t , x − g t , x j ∥ 2 ≤ ∫ x j x ∥ ∂ y g t , y ∥ 2 d y ≤ C ( t ∧ x − 1 / 2 ) ∫ x j x d y y ≤ C η ( t ∧ x − 1 / 2 ) . (140) \begin{aligned}
\lVert g_{t,x}-g_{t,x_j}\rVert_2\le\int_{x_j}^{x}\lVert\partial_y g_{t,y}\rVert_2\mathrm{d}y \\
&\le C(\sqrt{t}\wedge x^{-1/2})\int_{x_j}^{x}\frac{\mathrm{d}y}{y} \\
&\le C\eta(\sqrt{t}\wedge x^{-1/2}). \tag*{(140)}
\end{aligned} ∥ g t , x − g t , x j ∥ 2 ≤ ∫ x j x ∥ ∂ y g t , y ∥ 2 d y ≤ C ( t ∧ x − 1/2 ) ∫ x j x y d y ≤ C η ( t ∧ x − 1/2 ) . (140)
The weights are the positive masses
μ H ( [ x j , x j + 1 ] ) = 2 H β Γ ( β ) ( x j + 1 β − x j β ) = w j . \mu_H([x_j,x_{j+1}])=\frac{\sqrt{2H}}{\beta\Gamma(\beta)}(x_{j+1}^{\beta}-x_j^\beta)=w_j. μ H ([ x j , x j + 1 ]) = β Γ ( β ) 2 H ( x j + 1 β − x j β ) = w j .
Thus
∥ ∫ ε R g t , x μ H ( d x ) − ∑ j = 0 N − 1 w j g t , x j ∥ 2 ≤ ∑ j = 0 N − 1 ∫ x j x j + 1 ∥ g t , x − g t , x j ∥ 2 μ H ( d x ) ≤ C H η t H ≤ C H η L H . \begin{aligned}
\left\lVert\int_\varepsilon^R g_{t,x}\mu_H(\mathrm{d}x)-\sum_{j=0}^{N-1}w_jg_{t,x_j}\right\rVert_2
\le\sum_{j=0}^{N-1}\int_{x_j}^{x_{j+1}}\lVert g_{t,x}-g_{t,x_j}\rVert_2\mu_H(\mathrm{d}x) \\
&\le C_H\eta t^H\le C_H\eta L^H.
\end{aligned} ∫ ε R g t , x μ H ( d x ) − j = 0 ∑ N − 1 w j g t , x j 2 ≤ j = 0 ∑ N − 1 ∫ x j x j + 1 ∥ g t , x − g t , x j ∥ 2 μ H ( d x ) ≤ C H η t H ≤ C H η L H .
Adding the three errors, dividing by A H A_H A H , and applying the Wiener isometry proves (89) .
Resolution and value rate. For the parameters in (90) ,
L ε L β = R L − H = η L L H = L − γ . \sqrt{L}\,\varepsilon_L^\beta= R_L^{-H} = \eta_L L^H = L^{-\gamma}. L ε L β = R L − H = η L L H = L − γ .
A geometric partition, with a possibly shorter last interval, has
N L ≤ 1 + log ( R L / ε L ) log ( 1 + η L ) N_L \le1+\frac{\log(R_L/\varepsilon_L)}{\log(1+\eta_L)} N L ≤ 1 + log ( 1 + η L ) log ( R L / ε L )
≤ 1 + 2 L H + γ ( γ H + γ + 1 / 2 β ) log L = O H , γ ( L H + γ log L ) , \begin{aligned}
\le1+2L^{H+\gamma}\left(\frac{\gamma}{H}+\frac{\gamma+1/2}{\beta}\right)\log L \\
&=O_{H,\gamma}(L^{H+\gamma}\log L),
\end{aligned} ≤ 1 + 2 L H + γ ( H γ + β γ + 1/2 ) log L = O H , γ ( L H + γ log L ) ,
using log ( 1 + η ) ≥ η / 2 \log(1+\eta)\ge\eta/2 log ( 1 + η ) ≥ η /2 for 0 < η ≤ 1 0<\eta\le1 0 < η ≤ 1 . The path error bound gives
A N L , L / L = O ( L − 2 γ ) , B N L , L = O ( L − γ ) . A_{N_L,L}/L=O(L^{-2\gamma}),\qquad B_{N_L,L}=O(L^{-\gamma}). A N L , L / L = O ( L − 2 γ ) , B N L , L = O ( L − γ ) .
Substitution in (82) yields
∣ u N L ( L ) L − κ H ∣ ≤ C H , γ ( L − 1 + L − γ − 1 + L − 2 γ + L − γ ) ≤ C H , γ ( L − γ + L − 1 ) . \left|\frac{u_{N_L}(L)}{L}-\kappa_H\right|
\le C_{H,\gamma}\left(L^{-1}+L^{-\gamma-1}+L^{-2\gamma}+L^{-\gamma}\right)
\le C_{H,\gamma}(L^{-\gamma}+L^{-1}). L u N L ( L ) − κ H ≤ C H , γ ( L − 1 + L − γ − 1 + L − 2 γ + L − γ ) ≤ C H , γ ( L − γ + L − 1 ) .
A quadratic-hedging corollary
Let S S S be a continuous square-integrable F-martingale with
d ⟨ S ⟩ t = a t d t (141) \mathrm{d}\langle S\rangle_t=a_t\,\mathrm{d}t \tag*{(141)} d ⟨ S ⟩ t = a t d t (141)
where a , b a,b a , b satisfy definition 2.2. For ξ ∈ L 2 ( F T ) \xi\in L^2(\mathcal{F}_T) ξ ∈ L 2 ( F T ) , put ξ 0 = E [ ξ ∣ F 0 ] \xi_0=\mathbb{E}[\xi\mid\mathcal{F}_0] ξ 0 = E [ ξ ∣ F 0 ] . The Galtchouk–Kunita–Watanabe decomposition is
ξ = ξ 0 + ∫ 0 T Θ t d S t + L T , L 0 = 0 , ⟨ L , S ⟩ = 0 , (142) \xi=\xi_0+\int_0^T\Theta_t\,\mathrm{d}S_t+L_T,\qquad L_0=0,\qquad\langle L,S\rangle=0, \tag*{(142)} ξ = ξ 0 + ∫ 0 T Θ t d S t + L T , L 0 = 0 , ⟨ L , S ⟩ = 0 , (142)
where L L L is square-integrable and Θ \Theta Θ is a predictable integrand with E [ ∫ 0 T Θ t 2 d ⟨ S ⟩ t ] < ∞ \mathbb{E}\left[\int_0^T\Theta_t^2\,\mathrm{d}\langle S\rangle_t\right]<\infty E [ ∫ 0 T Θ t 2 d ⟨ S ⟩ t ] < ∞ , as in [30 ] . Assume this integrand has a continuous adapted version, which we fix. For initial cash c ∈ L 2 ( F 0 ) c\in L^2(\mathcal{F}_0) c ∈ L 2 ( F 0 ) and ϕ ∈ A T ( Θ ) \phi\in\mathcal{A}_T(\Theta) ϕ ∈ A T ( Θ ) , define
Q λ c : A T ( Θ ) ⟶ [ 0 , ∞ ) ϕ ⟼ 1 2 E [ ( ξ − c − ∫ 0 T ϕ t − d S t ) 2 ] + λ E [ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] . (143) \mathcal{Q}^{c}_{\lambda}:\quad
\begin{aligned}
\mathcal{A}_T(\Theta)&\longrightarrow[0,\infty)\\
\phi&\longmapsto\frac{1}{2}\mathbb{E}\left[\left(\xi-c-\int_0^T\phi_{t-}\,\mathrm{d}S_t\right)^2\right]+\lambda\mathbb{E}\left[\int_{[0,T]}b_t\,\mathrm{d}|\phi|_t\right].
\end{aligned} \tag*{(143)} Q λ c : A T ( Θ ) ϕ ⟶ [ 0 , ∞ ) ⟼ 2 1 E ( ξ − c − ∫ 0 T ϕ t − d S t ) 2 + λ E [ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] . (143)
Proposition J.1 (penalized mean-square identity). Under these assumptions,
inf ϕ ∈ A T Q λ c ( ϕ ) − 1 2 E [ ( ξ 0 − c ) 2 + L T 2 ] = V λ ( Θ ) . (144) \inf_{\phi\in\mathcal{A}_T}\mathcal{Q}^{c}_{\lambda}(\phi)-\frac{1}{2}\mathbb{E}\left[(\xi_0-c)^2+L_T^2\right]=V_\lambda(\Theta). \tag*{(144)} ϕ ∈ A T inf Q λ c ( ϕ ) − 2 1 E [ ( ξ 0 − c ) 2 + L T 2 ] = V λ ( Θ ) . (144) If a , b a,b a , b are continuous and Θ \Theta Θ satisfies assumption 6.1 in the same observation filtration, then
inf ϕ ∈ A T Q λ c ( ϕ ) − 1 2 E [ ( ξ 0 − c ) 2 + L T 2 ] ∼ λ → 0 κ H λ 2 H / ( 1 + H ) ∫ 0 T g H ( t ) d t . (145) \inf_{\phi\in\mathcal{A}_T}\mathcal{Q}^{c}_{\lambda}(\phi)-\frac{1}{2}\mathbb{E}\left[(\xi_0-c)^2+L_T^2\right]\mathop{\sim}_{\lambda\to0}\kappa_H\lambda^{2H/(1+H)}\int_0^T g_H(t)\,\mathrm{d}t. \tag*{(145)} ϕ ∈ A T inf Q λ c ( ϕ ) − 2 1 E [ ( ξ 0 − c ) 2 + L T 2 ] ∼ λ → 0 κ H λ 2 H / ( 1 + H ) ∫ 0 T g H ( t ) d t . (145) In particular, the second statement applies to the Volterra class of proposition 6.8 when its driving Brownian filtration is the observation filtration.
Proof. Using ( x + y ) 2 ≤ 2 x 2 + 2 y 2 (x+y)^2\le2x^2+2y^2 ( x + y ) 2 ≤ 2 x 2 + 2 y 2 , the admissibility and integrability assumptions give
E [ ∫ 0 T ϕ t 2 d ⟨ S ⟩ t ] ≤ 2 E [ ∫ 0 T ( ϕ t − Θ t ) 2 a t d t ] + 2 E [ ∫ 0 T Θ t 2 d ⟨ S ⟩ t ] < ∞ . \mathbb{E}\left[\int_0^T\phi_t^2\,\mathrm{d}\langle S\rangle_t\right]
\le2\mathbb{E}\left[\int_0^T(\phi_t-\Theta_t)^2a_t\,\mathrm{d}t\right]
+2\mathbb{E}\left[\int_0^T\Theta_t^2\,\mathrm{d}\langle S\rangle_t\right]<\infty. E [ ∫ 0 T ϕ t 2 d ⟨ S ⟩ t ] ≤ 2 E [ ∫ 0 T ( ϕ t − Θ t ) 2 a t d t ] + 2 E [ ∫ 0 T Θ t 2 d ⟨ S ⟩ t ] < ∞.
A càdlàg finite-variation path has countably many jumps. Since d ⟨ S ⟩ t = a t d t \mathrm{d}\langle S\rangle_t=a_t\,\mathrm{d}t d ⟨ S ⟩ t = a t d t , replacing ϕ t \phi_t ϕ t by its predictable left-limit process ϕ t − \phi_{t-} ϕ t − leaves this integral unchanged. Define the square-integrable martingale
N t : = ∫ 0 t ( Θ r − ϕ r − ) d S r , N 0 = 0. N_t:=\int_0^t(\Theta_r-\phi_{r-})\,\mathrm{d}S_r,\qquad N_0=0. N t := ∫ 0 t ( Θ r − ϕ r − ) d S r , N 0 = 0.
The hedging error decomposes as
ξ − c − ∫ 0 T ϕ t d S t = ( ξ 0 − c ) + L T + N T . \xi-c-\int_0^T \phi_t\,\mathrm{d}S_t=(\xi_0-c)+L_T+N_T. ξ − c − ∫ 0 T ϕ t d S t = ( ξ 0 − c ) + L T + N T .
These three terms are orthogonal in L 2 ( Ω ) L^2(\Omega) L 2 ( Ω ) : the martingale increments have zero conditional mean given F 0 \mathcal{F}_0 F 0 , while ⟨ L , S ⟩ = 0 \langle L,S\rangle=0 ⟨ L , S ⟩ = 0 gives the last equality below. First,
E [ ( ξ 0 − c ) N T ] = E [ ( ξ 0 − c ) E [ N T ∣ F 0 ] ] = 0. \mathbb{E}[(\xi_0-c)N_T]=\mathbb{E}[(\xi_0-c)\mathbb{E}[N_T\mid\mathcal{F}_0]]=0. E [( ξ 0 − c ) N T ] = E [( ξ 0 − c ) E [ N T ∣ F 0 ]] = 0.
Similarly,
E [ ( ξ 0 − c ) L T ] = E [ ( ξ 0 − c ) E [ L T ∣ F 0 ] ] = 0. \mathbb{E}[(\xi_0-c)L_T]=\mathbb{E}[(\xi_0-c)\mathbb{E}[L_T\mid\mathcal{F}_0]]=0. E [( ξ 0 − c ) L T ] = E [( ξ 0 − c ) E [ L T ∣ F 0 ]] = 0.
Finally,
E [ L T N T ] = E [ ⟨ L , N ⟩ T ] = E [ ∫ 0 T ( Θ t − ϕ t − ) d ⟨ L , S ⟩ t ] = 0. \mathbb{E}[L_TN_T]=\mathbb{E}[\langle L,N\rangle_T]=\mathbb{E}\left[\int_0^T(\Theta_t-\phi_{t-})\,\mathrm{d}\langle L,S\rangle_t\right]=0. E [ L T N T ] = E [⟨ L , N ⟩ T ] = E [ ∫ 0 T ( Θ t − ϕ t − ) d ⟨ L , S ⟩ t ] = 0.
Itô’s isometry gives
E [ N T 2 ] = E [ ∫ 0 T ( Θ t − ϕ t − ) 2 d ⟨ S ⟩ t ] = E [ ∫ 0 T a t ( Θ t − ϕ t ) 2 d t ] . \mathbb{E}[N_T^2]=\mathbb{E}\left[\int_0^T(\Theta_t-\phi_{t-})^2\,\mathrm{d}\langle S\rangle_t\right]=\mathbb{E}\left[\int_0^T a_t(\Theta_t-\phi_t)^2\,\mathrm{d}t\right]. E [ N T 2 ] = E [ ∫ 0 T ( Θ t − ϕ t − ) 2 d ⟨ S ⟩ t ] = E [ ∫ 0 T a t ( Θ t − ϕ t ) 2 d t ] .
Consequently,
Q λ ( ϕ ) = 1 2 E [ ( ξ 0 − c ) 2 + L T 2 ] + 1 2 E [ N T 2 ] + λ E [ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] = 1 2 E [ ( ξ 0 − c ) 2 + L T 2 ] + J λ ( ϕ ; Θ ) . \begin{aligned}
\mathcal{Q}_{\lambda}(\phi)=\frac{1}{2}\mathbb{E}[(\xi_0-c)^2+L_T^2]+\frac{1}{2}\mathbb{E}[N_T^2]+\lambda\mathbb{E}\left[\int_{[0,T]}b_t\,\mathrm{d}|\phi|_t\right] \\
&=\frac{1}{2}\mathbb{E}[(\xi_0-c)^2+L_T^2]+J_{\lambda}(\phi;\Theta).
\end{aligned} Q λ ( ϕ ) = 2 1 E [( ξ 0 − c ) 2 + L T 2 ] + 2 1 E [ N T 2 ] + λ E [ ∫ [ 0 , T ] b t d ∣ ϕ ∣ t ] = 2 1 E [( ξ 0 − c ) 2 + L T 2 ] + J λ ( ϕ ; Θ ) .
Take infima to obtain (144) , then apply theorem 6.2 for the asymptotic formula.
Corollary J.2 (initial hedge position). If E [ ∣ Θ 0 ∣ ] < ∞ \mathbb{E}[|\Theta_0|]<\infty E [ ∣ Θ 0 ∣ ] < ∞ , replacing the pre-initial hedge position Θ 0 \Theta_0 Θ 0 by a fixed x 0 ∈ R x_0\in\mathbb{R} x 0 ∈ R changes the value by at most λ b + E [ ∣ x 0 − Θ 0 ∣ ] \lambda b_+ \mathbb{E}[|x_0-\Theta_0|] λ b + E [ ∣ x 0 − Θ 0 ∣ ] . Thus (145) is unchanged.
Proof. Apply proposition 2.3 to (144) .
AI Assistance Declaration
During the preparation of this manuscript, the author used AI tools for language editing and manuscript organization. The author takes responsibility for the mathematical statements, proofs, computations, references, and final content.
References