Pathwise Uniqueness for Multiplicative Young and Rough Differential Equations Driven by Fractional Brownian Motion

See Thm. 3.8 for the case H \in (1/2,1) and Thm. 4.10 for H \in (1/3,1/2].

This follows from a genelarised Yamada-Watanabe theorem of Kurtz [kurtz07]. One can consult [matsuda22, Appendix B] for more detailed discussion.

The proof is given in App, ?.

For notational concision, we write Y^s \vcentcolon = Y^{H, s}. Since (s, \infty) \ni t \mapsto K_H(t, r) is smooth for any r \in [0, s], the process (s,\infty) \ni t\mapsto Y^s_t is smooth, hence we can legitimately take its derivative in this region, which is all that the lemma requires. The estimate on Y^{s}_{u, t} readily follows from that on \dot{Y}^s_t, through the identity Y^{s}_{u, t} = \int_u^t \dot{Y}^s_r \, \mathrm{d} r. Therefore, we focus on the estimate on \dot{Y}^s_t.

To this end, we use the Mandelbrot–Van Ness representation [mandelbrot68]:

where \bar{W} is a two-sided Brownian motion and \bar{c}_H is a suitably chosen constant. We set

and \bar{Y}^s_t \vcentcolon = \mathbb{E}[B^H_t \vert \bar{\mathcal{F}}_s]. Since \mathcal{F}_t \subseteq \bar{\mathcal{F}}_t and Y^s_t = \mathbb{E}[\bar{Y}^s_t \vert \bar{\mathcal{F}}_s], we have \dot{Y}^s_t = \mathbb{E}[\dot{\bar{Y}}^s_t \vert \bar{\mathcal{F}}_s]. Therefore, it suffices to estimate \dot{\bar{Y}}^s_t, and we observe that

The proof is given in App. ?.

Since \mathbb{B}^{H, s} satisfies Chen’s relation (?) and Y^{H, s} is smooth, it is easy to see that \tilde{\mathbb{B}}^{H, s} also satisfies Chen’s relation. To see the regularity estimate, by Kolmogorov’s continuity criterion for rough paths [FH20, Thm. 3.1], it suffices to show that

But, since we already know the estimates on \mathbb{B}^{H, s} and on \dot{Y}^{H, s} from Lem. 2.11, the bound (Proof) easily follows.

To see that (\tilde{B}^{H, s}, \tilde{\mathbb{B}}^{H, s}) is a geometric rough path of order \alpha, by [FH20, Prop. 2.8] it suffices to show that the rough path is weakly geometric, i.e.

but this algebraic identity is easy to verify, since (B^{H}, \mathbb{B}^H) is geometric.

The proof is given in App. ?.

For notational concision we write \tilde{B}^s \vcentcolon = \tilde{B}^{H, s}. We first note that, since \tilde{B}^s is Gaussian, the process (recall \rho_H from (?))

is orthogonal to \tilde{B}^{s, i} in L^2_{\omega}. Therefore, also applying Lem. B.3, we have

and we obtain

Gaussian integration by parts yields

Combining (?) and (?), we obtain the desired identity.

For simplicity, in the remainder of the proof we write B \vcentcolon = B^H and do not keep track of dependencies on \beta, l, p, \tilde{\gamma}, T.

Our strategy of proof is to apply Gerencsér’s shifted stochastic sewing, [gerencser2022, Lem. 2.2] (equivalently Lem. 2.4 with \alpha = 0 and v = s - (t-s)) to the germ

As we demonstrate below, the shifted conditional, expectation and coefficients, along with the shifted stochastic sewing lemma, allow us to obtain the claimed estimate. We note that the shift is necessary, for instance to bound (?) below.

We decompose the proof into six steps. First demonstrate that the integral \int_s^t f(X_r)\,\mathrm{d} B^H_r can be realised as the (shifted) stochastic sewing of (Proof).

Step 1.

This is equivalent to the claim that the following limit holds in L^p(\Omega;\mathbb{R}),

Namely, this shows that the germ (A_{s, t}) provides a correct approximation to the integral, (?). To prove this identity, we first set v \vcentcolon = s - (t-s) and define the auxiliary germ

It is directly clear that the Riemann sums of A^0 and A coincide. To wit, appealing to the uniqueness element of the stochastic sewing lemma, as in Rem. 2.7, it suffices to check that

and

The latter is immediate since \mathbb{E}_{s-(t-s)} \left[A_{s,t}-A^0_{s,t}\right] =0. The former also follows rather straightforwardly, we apply the triangle inequality, the fundamental estimate of Young integration (?) and the assumptions on f,\, (X,X') to find,

Hence, (?) follows since here we only consider H>1/2. Therefore, in order to obtain (?) it is enough to show that

We achieve this by applying the vanilla sewing lemma, (Lem. 2.3) with E=L^p(\Omega;\mathbb{R}^{d_2}).

Indeed, with v = s - (t-s), we have

By the remainder estimate of Young integration (?),

we obtain

In addition, since

where R is defined by (?), we get

Combining these estimates, we obtain

As (1+\beta) H > 1, appealing this time to the uniqueness element of the deterministic sewing lemma (see Rem. 2.7), the identity (?) is established.

We are now in a position to exploit our choice of germ, given by (Proof) to obtain the claimed estimate. We first recall the Volterra representation of the fBm, given by (?); i.e. for v \geq 0 we set Y^v \vcentcolon = Y^{H, v} and \tilde{B}^v \vcentcolon = \tilde{B}^{H, v} where the right hand sides are given by (?) and (?) respectively.

By Lem. 2.13, given any matrix a \in \mathcal{M}(d_1,d_2) and v \leq s \leq t, g \in C^1(\mathbb{R}^{d_2}, \mathbb{R}) it holds that

with \rho^2(v,r) \vcentcolon = \rho_H^2(v, r) defined by (?). We also note that

Recall the germ A_{s,t} from (Proof) we are now in a position to check the analytic bounds (?). The first is checked by showing that

which is done in Step 2 below. The second estimate of (?) is shown through Steps 2 - 5.

Step 2.

For v \vcentcolon = s - (t-s), by (?) and the trivial fact that X_{v},\,X'_v are measurable with respect to \mathcal{F}_v we observe that

Since x' \tilde{B}^v_r is a centred Gaussian random variable with variance \rho^2(v, r) x' (x')^{\mathrm{T}}, we find

and by Lem. A.1 and (?)

By Lem. 2.11 we have

\norm{\dot{Y}^v_r}_{L^p_{\omega}} \lesssim_p (r-v)^{H-1}.

Therefore,

To estimate I_2, we note that by (?) which is obtained from our averaging result Lem. 2.13, applied with g(z) = f(x + x'y_{v, r} + x' z), we find

Then, appealing to Lem. A.1 we obtain a regularising effect by the estimate,

so that making use of (?) and (?), we obtain

This concludes Step 2 and the proof of (?).

In the remaining steps we show that for u \in (s, t) with

we have

which is the final estimate needed to show that A_{s,t} satisfies the analytic bounds of the stochastic sewing lemma, (?). We recall that imposing the additional condition of (?) does not alter the applicability of the stochastic sewing lemma, see Rem. 2.6.

Step 3.

We first perform some simple algebraic manipulations. Let us set

It is clear that,

One may then check that \mathbb{E}_{v_1} \left[\delta A_{s, u, t}\right] = \mathbb{E}_{v_1}[I_3 + I_4], where

Since the estimate for I_3 and I_4 are similar (the latter requires exchanging (s, u) for (u, t)), we only present the estimates for I_3. Given j=1,2, we observe

By (?), the first term is equal to

and by (?) the second term is equal to

Hence, we write I_3 = I_{3,1} + I_{3,2}, where

Step 4. In view of Step 3, in order to prove (?), it remains to show

along with similar arguments for I_4 which we do not present.

In this step, we estimate I_{3,1}.

We decompose I_{3,1} = I_{3,1,1} + I_{3,1,2}, where

and

To estimate I_{3,1,1}, we apply Lem. A.1 to obtain the Lipschitz estimate

which yields

Recalling the definition of the remainder R of a controlled path from (?), we observe that X_{v_1} + X'_{v_1}B_{v_1, v_2} - X_{v_2} = - R_{v_1, v_2}. Therefore,

By Hölder’s inequality and recalling the norm \@ifstar \left|\mkern-1.5mu\left|\mkern-1.5mu\left| \mathopen{|\mkern-1.5mu|\mkern-1.5mu|} \right \mathclose{|\mkern-1.5mu|\mkern-1.5mu|} |\mkern-1.5mu\right|\mkern-1.5mu\right| {X}_{p, \beta}, defined by (?), we find

By Lem. 2.11 we have \norm{\dot{Y}^{v_2}_r}_{L^{4p}_{\omega}} \lesssim (r-v_2)^{H-1} and \norm{Y_{v_2, r}^{v_2}}_{L^{4p}_{\omega}} \lesssim (r-v_2)^H. Hence, also using the bound (?) to estimate \rho(v_2, r), we have

and

Recalling v_1, v_2 from (?) and the condition (?) on u, we obtain

an estimate compatible with (?).

We move to estimate I_{3,1,2}. By Lem. A.2,

Thus, by Hölder’s inequality,

Recalling the definition of \@ifstar \left|\mkern-1.5mu\left|\mkern-1.5mu\left| \mathopen{|\mkern-1.5mu|\mkern-1.5mu|} \right \mathclose{|\mkern-1.5mu|\mkern-1.5mu|} |\mkern-1.5mu\right|\mkern-1.5mu\right| {X}_{2p, \beta} from (?), applying Lem. 2.11 and (?) we then find

which, combined with the estimate of I_{3,1,1} in the previous paragraph, yields (?).

Step 5. In this last step, we estimate I_{3,2} to obtain (?). By adding and subtracting, we write I_{3,2} = I_{3,2,1} + I_{3,2,2} + I_{3,2,3}, where

and

The estimate of I_{3,2,1} and that of I_{3,2,2} are similar to that of I_{3,1,1} and that of I_{3,1,2} respectively; we only demonstrate how to estimate I_{3,2,1}. By Lem. A.1,

Therefore,

As above, recalling the definition of \@ifstar \left|\mkern-1.5mu\left|\mkern-1.5mu\left| \mathopen{|\mkern-1.5mu|\mkern-1.5mu|} \right \mathclose{|\mkern-1.5mu|\mkern-1.5mu|} |\mkern-1.5mu\right|\mkern-1.5mu\right| {X}_{p, \beta}, the estimate on Y^{v_2} from Lem. 2.11 and the bound (?), we get

Applying the estimate (?) for \rho(v_2, r), we get

We finally estimate I_{3,2,3}. Using the bounds (?) and (?),

Therefore,

yielding the estimate (?), which in turn gives us (?), completing the proof.

Let \alpha \in (1/2,H) be arbitrarily close to H and (X, B^H) and (Y, B^H) be two weak solutions defined on the same filtered probability space (\Omega,\mathcal{F},(\mathcal{F}_t),\mathbb{P}) and driven by the same (\mathcal{F}_t)-fBm, B^H. Since the requirement on \gamma \in (\frac{1}{2H},1) is given by an open interval it is always possible to find a \gamma' \in (\frac{1}{2H},\gamma). We then let (\sigma_n)_{n=1}^{\infty} be a smooth approximation to \sigma converging in \mathcal{C}^{\gamma'}(\mathbb{R}^{d_1};\mathcal{M}(d_1,d_2)) (which exists by the compact embedding between Hölder–Besov spaces).

The remainder estimate of Young integration (?) gives

for Z \in \{X, Y\}.

From the estimate (?) and convergence of the \sigma_n in \mathcal{C}^{\gamma'}(\mathbb{R}^{d_1};\mathcal{M}(d_1,d_2)), it follows that

uniformly in [0, T]. As the computation (?) shows, we have

where

Hence, we obtain

Now our task is to prove the convergence of V_n. By Lem. 3.9, for (s,t) \in [0,T ]^2_{\leq} we have

with

One technical problem is that the Gubinelli derivative \theta \sigma(X_s) + (1-\theta) \sigma(Y_s) might be degenerate, that is, the matrix

could have an eigenvalue 0. (In the end, we will show that X = Y, hence it must be non-degenerate; but a priori we do not know this.) Because of this technical problem, we proceed as follows.

Step 1. We first assume that

where l is the assumed ellipticity constant of \sigma. This condition implies that for all x,y\in \mathbb{R}^{d_1} and \theta \in (0,1) we have

Therefore, in combination of Lem. 3.9, we see that

Prop. 3.7 and standard embedding of Hölder–Besov spaces, yield

By Kolmogorov’s continuity theorem, we see that there exists a process V^k such that for any \beta < \gamma H,

Since \gamma' H > 1/2, we may take \beta > 1/2. Therefore, recalling (?) we observe

As this shows that X - Y solves the linear Young differential equation

we see that X - Y = 0 a.s.

Step 2. Now we do not assume (?). The new ingredient is a stopping time argument. We choose \varepsilon > 0 so that

and we set T^{(0)} \vcentcolon = 0 and inductively

Using the notation \llbracket X \rrbracket_{\mathcal{C}^{\alpha}_T} from (?), if T^{(i)} \leq T,

Hence, the a priori estimate (?) implies that

uniformly over i, as long as T^{(i)} \leq T.

To see that X = Y a.s. up to time T^{(1)}, let \sigma^{(1)} be a \gamma-Hölder map such that \sigma^{(1)} = \sigma in an \varepsilon-neighbourhood of the initial condition x and such that

We set

and

Up to time T^{(1)}, we have

Due to our choice of \sigma^{(1)}, the argument of Step 1 shows that V^{(1), k}_n converges to some V^{(1), k} in \mathcal{C}_T^{\beta}, and that X = Y up to time T^{(1)}.

Step 3. Obviously we want to repeat the operation of Step 2, but now there is a small problem that X_{T^{(1)}} is random. For this sake, let (x_m)_{m \in \mathbb{N}} be a countable dense set of \mathbb{R}^{d_1}, and let \sigma^m be a \gamma-Hölder map such that \sigma^m = \sigma in an \varepsilon-neighbourhood of x_m and such that

For each m and Z \in \{X, Y\}, we define the approximation,

along with the cut-off equations,

and approximate linearisation,

Notice that X^{(2, m)}, Y^{(2, m)} are both adapted. We claim that, for Z \in \{X, Y\},

Indeed, by the remainder estimate of Young’s integral (?), the estimate (?) is obvious if t \leq T^{(1)} or if s \geq T^{(1)}. Suppose that s < T^{(1)} < t. If \abs{x_m - X_{T^{(1)}}} \leq \varepsilon/2, then

Since \abs{x_m - X_{T^{(1)}}} = \abs{x_m - Y_{T^{(1)}}} \leq \varepsilon/2, we have

and the estimate (?) follows. The case where \abs{x_m - X_{T^{(1)}}} > \varepsilon/2 is similar.

Now we check that the Gubinelli derivative \Sigma^{\theta, m} \vcentcolon = \theta \Sigma^{X,m} + (1-\theta) \Sigma^{Y, m} is non-degenerate. Indeed, for t \leq T^{(1)} we have

and for t > T^{(1)} the condition (?) implies that \Sigma^{\theta, m}(t) \Sigma^{\theta, m}(t)^{\top} \geq l^{-1}/4. Therefore, Prop. 3.7 and the Kolmogorov continuity theorem show that V^{(2, m)}_n converges to some V^{(2, m)} in L^p_{\omega}. By the diagonalization argument, we may suppose that almost surely for every m \in \mathbb{N} and \delta \in (0, \gamma H) we have

Since (x_m) is dense, we can find a random m = m(\omega) so that \abs{x_m - X_{T^{(1)}}} < \varepsilon/2. We then have

up to t \leq T^{(2)}, hence X = Y on [0, T^{(2)}]. It is now clear that this algorithm can be continued, and at some point we must have T^{(i)} \geq T due to (?).

Let \alpha < H, but sufficiently close to H. We recall the seminorm \llbracket \,\cdot\, \rrbracket_{\mathcal{C}^{\alpha}_{[s, t]}} from (?). Since

the remainder estimate of Young integration, estimate (?), gives

By [friz_victoir_2010, Lem. 10.7], we have

Now, due to Gaussianity of B^H and as a consequence of the Kolmogorov continuity criterion, for all p < \infty, there exists a C\coloneqq C(p, \alpha, T, \norm{\sigma}_{\mathcal{C}_x^{\gamma}})>0 such that

Combining (Proof) and (?), we get

\norm{X_{s,t}}_{L^p_\omega} \lesssim (t-s)^H.

In particular, we have \norm{\sigma(X)_{s, t}}_{L^p_\omega} \lesssim (t-s)^{\gamma H}. To get an estimate of the remainder R_{s,t} = X_{s, t} - \sigma(X_s) B_{s, t}^H, it suffices to apply the sewing lemma in L^p_{\omega} with germ A_{s, t} \vcentcolon = \sigma(X_s) B_{s, t}^H,

since

See App. ?.

In view of Lem. 2.4 and Lem. B.6, it suffices to show

where

A_{s, t} \vcentcolon = 2^{-1}(F(\xi, B^H_{\tau, s}) + F(\xi, B^H_{\tau, t})) \tilde{B}^{H, \tau}_{s, t}.

But the proof goes in the same way as Lem. B.3, so we omit the proof.

For concision, throughout the proof we simply write B \vcentcolon = B^H. Let us set

The integral is understood as rough integral: the path

r \mapsto f(X_s + X_s' B_{s, r})

is controlled by B with Gubinelli derivative \nabla f(X_s + X_s' B_{s, r}) X'_s. In particular, the estimate (?) yields

for \alpha \in ( 1/3, H). Hence, the uniqueness part of the sewing lemma (see Rem. 2.7) implies that

where \pi is a partition of [s, t]. As in Prop. 3.7, we apply the stochastic sewing to A_{s, t}. This time, Lê’s original version [le20] is sufficient (that is \alpha = 0 and v = s in Lem. 2.4). Precisely, we prove the estimates

for s < u < t with appropriate \varepsilon_1, \varepsilon_2 > 0 and proportional constants \Gamma_1, \Gamma_2.

We split the rest of the proof into three steps.

Step 1. We estimate \norm{\delta A_{s, u, t}}_{L^p_{\omega}}. Writing Y \vcentcolon = Y^{H, u} and \tilde{B} \vcentcolon = \tilde{B}^{H, u} as in (?) and (?) respectively, we observe

The integral (?), by the triangle inequality, is bounded by

Defining the remainder R as in Def. 3.6, we have

and since, due to Lem. 2.11 we have \norm{\dot{Y}_r}_{L^{2p}_{\omega}} \lesssim_{p} (r - u)^{H-1}, it follows that

To estimate (?), for v \in \{s, u\} we apply Lem. 4.6 with

to obtain the estimate

Since

we see that the L^p_{\omega}-norm of the integral (?) is bounded by (up to constant)

Hence,

Step 2. Next we estimate \mathbb{E}[\delta A_{s, u, t} \vert \mathcal{F}_s], which is essentially done in Prop. 3.7. We define Y and \tilde{B} as in Step 1, and we write \mathbb{E}_u[\,\cdot\,] \vcentcolon = \mathbb{E}[\,\cdot\, \vert \mathcal{F}_u]. We have

For v \in \{s, u\}, we observe

with \rho^2(u, r) \vcentcolon = \rho^2_H(u, r) defined by (?). As in the proof of Prop. 3.7 (especially (?)), Lem. 2.13 implies that

Hence, \mathbb{E}_u \left[\delta A_{s, u, t}\right] = J_1 + J_2, where

Then the rest of the argument is identical to the proof of Prop. 3.7. (J_1 corresponds to I_{3,1} and J_2 to I_{3,2}.) In particular, we obtain

Step 3. Thanks to the estimates (?) and (?), we are in the setting of Lem. 2.4 with \alpha = 0 and v=s. Therefore, the remainder estimate of Lem. 2.4 yields

On the other hand, we have

By Lem. 2.11, the first term is bounded by \norm{f}_{\mathcal{C}^{\tilde{\gamma}}_x} (t-s)^{2H} up to constant, while by Lem. 4.6 the second term is bounded by \norm{f}_{\mathcal{C}^{\tilde{\gamma}}_x} (t-s)^{(1+\tilde{\gamma})H} up to constant. Therefore, since \tilde{\gamma}<1 and |t-s|\leq 1,

Finally, we observe

Combining (?), (?) and (?), we obtain the desired estimate.

For concision, we write B \vcentcolon = B^H and \mathbb{B} \vcentcolon = \mathbb{B}^H. We fix a \tau \leq s, and we set

Step 1. We show that \bar{A}_{s, t} is a correct approximation to the iterated integral \int_s^t f(X_r) B_{\tau, r}^i \, \mathrm{d} B_r^j. For this sake, we set

By definition of the rough integral, we have

In view of the uniqueness part of the sewing lemma (see Rem. 2.7), our goal is to prove \abs{\bar{A}_{s, t} - \bar{A}'_{s, t}} \lesssim_{f, X, X', \omega} (t-s)^{3H}. We compute

By the remainder estimate of the rough integral (?), we have

whence \bar{A} is a correct approximation.

Step 2. This time we do not need the stochastic sewing; the deterministic sewing (Lem. 2.3) in L^p_{\omega} suffices. We compute

Using Chen’s relation

\mathbb{B}^{i, j}_{u, t} = \mathbb{B}^{i, j}_{s, t} - \mathbb{B}^{i, j}_{s, u} - B^i_{s, u} B^j_{u, t},

we obtain

Since

and by Prop. 4.7

using Cauchy–Schwarz inequality we obtain

As (2 + \tilde{\gamma}) H > 1, the sewing lemma in L^p_{\omega} gives

It remains to set \tau = s.

For concision, throughout the proof we write B \vcentcolon = B^H and \mathbb{B} \vcentcolon = \mathbb{B}^H. We fix a \tau \in [0, s]. To estimate the iterated integral

our choice of the germ is

Step 1. To see that \tilde{A}_{s, t} is a correct approximation, we set

By definition of the rough integral, we have

In view of the uniqueness part of the sewing lemma (see Rem. 2.7), our goal is to prove \abs{\tilde{A}_{s, t} - \tilde{A}'_{s, t}} \lesssim (t-s)^{3H}. We observe

By the remainder estimate of the rough integral (?), we have

Hence, \tilde{A}_{s, t} is a correct approximation.

Step 2. Again we do not need the stochastic sewing; the deterministic sewing lemma (Lem. 2.3) in L^p_{\omega} suffices. We compute

By Prop. 4.7 and Prop. 4.8,

Since (2+\tilde{\gamma})H > 1, we are in the setting of the sewing lemma (Lem. 2.3). The remainder estimate yields

Setting \tau = s, we get

By Prop. 4.8,

and the proof is complete.

We assume that (?) holds, as the general case is handled by stopping time arguments as exhibited in the proof of Thm. 3.8. Let (X, B^H) and (Y, B^H) be two weak solutions to (?) defined on the same filtered probability space (\Omega,\mathcal{F},(\mathcal{F}_t)_{t\in [0,T]},\mathbb{P}) and driven by the same (\mathcal{F}_t)-fBm, B^H. Since the requirement on \gamma \in ((1-H/H,2)) is given by an open interval it is always possible to find a \gamma' \in ((1-H)/H,\gamma). We then let (\sigma_n)_{n=1}^{\infty} be a smooth approximation to \sigma converging in \mathcal{C}^{\gamma'}(\mathbb{R}^{d_1};\mathcal{M}(d_1,d_2)) (which exists by the compact embedding between Hölder–Besov spaces). By [FH20, Lem 7.2] or [dareiotisgerencser22, Eq. (1.8)], we have

and similarly for Y. In view of (?), we then have

and so by Lem. 4.11,

By repeating the argument of Lem. 3.9, we can show that (X, \sigma(X)) \in \mathfrak{D}^{1, l}_{B^H} and similarly for Y. Furthermore, by the condition (?), as checked in Step 1 of the proof of Thm. 3.8, we have

Therefore, by Prop. 4.7 & 4.8 (with f = \partial_k \sigma_n - \partial_k \sigma_m),

and the Kolmogorov criterion for rough paths ([FH20, Thm. 3.1]), the sequence of the lifts \mathbf{G}^k[\sigma_n] converges in L^p(\Omega;\mathbb{R}) to some limit \mathbf{G}^k[\sigma] = (G^k[\sigma], \mathcal{G}^k[\sigma]) in the metric d_{\alpha, \gamma'} defined by (?). Namely, for all k = 1, 2, \ldots, d_1 and p \in [1, \infty),

The stability estimate (?) yields

The remainder estimate (?) shows that X-Y is controlled by (G^k)_{k=1}^{d_1} with Gubinelli derivative X - Y. Moreover, similarly to \mathbf{G}^k[\sigma], by Prop. 4.9 (with f = \partial_k \sigma_n - \partial_k \sigma_m, g = \partial_l \sigma_n and f = \partial_k \sigma_n, g = \partial_l \sigma_n - \partial_l \sigma_m) and the Kolmogorov criterion for rough paths, the sequence of the rough paths

converges to some limit \hat{\mathbf{G}}[\sigma]. We claim that

We note that the integral in the right-hand side is defined as rough integral with respect to \hat{\mathbf{G}}[\sigma], which is well-defined since X-Y is controlled by (G^k)_{k=1}^{d_1} with Gubinelli derivative X-Y. To see (?), we again rely on the uniqueness part of the sewing lemma (Rem. 2.7). By definition we have

where

By Prop. 4.8 & 4.9,

and as (1+\gamma)H > 1 the identity (?) is established.

Hence, X-Y solves the linear rough differential equation

and its uniqueness implies that for \mathbb{P}-a.e. \omega we have X(\omega) = Y(\omega), which concludes the proof.

The proof is completely pathwise. We set

where as usual X^k denotes the k^{\text{th}} element of X \in \mathbb{R}^{d_1} and (X')^k denotes the k^{\text{th}} row of the matrix X' \in \mathcal{M}(d_1,d_2), similarly for Y,\,Y'.

In addition, the second term in A^1_{s, t} should be understood as a d_1-dimensional vector whose i^{\text{th}} component is given by

Note that

Therefore, by the uniqueness part of the deterministic sewing lemma (Rem. 2.7), it suffices to show \abs{A^1_{s, t} - A^2_{s, t}} \lesssim (t-s)^{3H}. It follows from the mean value theorem that we have both

and

By (?) and (?), we compute

where

and

By the remainder estimate of the rough integral (?) and assumed regularity of \rho, both \Delta^{1, k}_{s, t} and \Delta^{2, k}_{s, t} are bounded by (t-s)^{3H} up to constant,

and the proof is complete.

We write

Then \mathcal{P}_{t\Gamma} f = \Lambda_{\Gamma^{-1}} \mathcal{P}_{t} \Lambda_{\Gamma} f. Since

we can assume that \Gamma = I_d. Then the claim follows from [dareiotis_gerenscer_le_23_euler, Lem. 2.3], note that the inclusion of t=1 in the definition of the inhomogeneous Hölder-Besov norm makes no difference to the arguments given therein..

We first claim

Indeed, we have

By [dareiotis_gerenscer_le_23_euler, Prop. 2.7],

which yields the estimate (Proof).

Next we prove the estimate (?) for \alpha = \beta \notin \mathbb{N}. The proof uses the characterization of Besov spaces by Littlewood–Paley blocks [bahouri_chemin_danchin_11_fourier]. The case for \alpha = \beta = 0 is proved by (Proof). Let \{\Delta_j\}_{j=-1}^{\infty} be Littlewood–Paley blocks. Since \Delta_j and \mathcal{P}_{\Gamma_i t} commute, the estimate (Proof) yields

Therefore,

Next we prove the estimate (?) for \alpha = \beta \in \mathbb{N}. Since we only need the case \alpha = \beta =1, we only prove that case. We have

By (?),

which is then bounded by 2 \abs{\Gamma_1 - \Gamma_2} \norm{f}_{C^1} \abs{x - y}.

Finally, we prove the general case. Since

we have

As for the first term, the estimate (?) with \alpha = \beta and Lem. A.1 imply

The estimate of the second term is similar.

For notational concision we write \tilde{B} \vcentcolon = \tilde{B}^{H, s} and \tilde{\mathbb{B}} \vcentcolon = \tilde{\mathbb{B}}^{H, s}. If i = j, the identity is trivial, since the rough path of \tilde{B} is geometric. We therefore assume below that i \neq j.

Since i \neq j, (B^{H, i}, Y^{H, s, i}) and (B^{H, j}, Y^{H, s, j}) are independent, and the random variable \tilde{\mathbb{B}}^{H, s, i, j} defined by (?) belongs to the second order Wiener chaos. (For the basic properties of the Wiener chaos, see [nualart06, Chapter 1.1].) The orthogonality in the Wiener chaos implies that

Since (\tilde{B}^i, \tilde{B}^j) \overset{\operatorname{d}}{=} (\tilde{B}^j, \tilde{B}^i), we have

where we use the fact that (\tilde{B}, \tilde{\mathbb{B}}) is geometric. Therefore,

and the proof is complete.

For simplicity, we drop scripts on H. Recalling the Volterra representation (?), we set

Note that Y^{\tau, v} = \mathbb{E}[\tilde{B}^{\tau} \vert \mathcal{F}_v] and \tilde{B}^{\tau} - Y^{\tau, v} = \tilde{B}^{v} is independent of \mathcal{F}_v. We have

To estimate the first three terms, observe by Lem. 2.11 that

and by (?) that \norm{\tilde{B}^v_{s, t}}_{L^p_{\omega}} \lesssim_p (t-s)^H. Hence, we obtain

Similarly, A^2_{s,t} equals to

and

To complete the proof, it remains to observe

which follows from Lem. B.1.

For simplicity, we drop scripts on H. When H > 1/2, the claim follows from the remainder estimate of the Young integral (?). Hence, we assume that H \leq 1/2.

We set

The integral \int_s^t g(\tilde{B}^{\tau}_r) \, \mathrm{d} \tilde{B}^{\tau}_r is understood as rough integral:

The proof relies on the uniqueness part of the deterministic and stochastic sewing lemma (Rem. 2.7). We have

where

\abs{R_{s, t}} \lesssim_{\norm{g}_{C^2_x}} \abs{\tilde{B}^{\tau}_{s, t}}^3.

Setting A^3_{s, t} \vcentcolon = A^2_{s, t} - A^1_{s,t} - R_{s, t}, we therefore see that

To estimate the right-hand side, we will apply the uniqueness part of Lem. 2.4. Using (?) and Lem. B.2, we obtain

Hence, applying the uniqueness part of Lem. 2.4, the claim is proven in view of (?) and (?).

Recalling (?) and observing

it suffices to prove

The estimate (Proof) is straightforward. Indeed, noting that r/t \leq 1, r/s \leq 1 and s-r \leq t-r, the left-hand side is bounded by

(s-v)^{2H - 1} up to constant.

To prove (?), using again r/t \leq 1 and also for r \in [0, v]

the left-hand side of (?) is bounded by (s-v)^{2H-1} I, where

Therefore, it suffices to show I \lesssim 1.

Applying the bound

we get

When H = 1/2 the claim is trivial, so we suppose that H < 1/2.

Proof of 1. By (?), we have

\mathbb{E}[\abs{B^{H, v}_{s, t}}^2] \lesssim (t-s)^{2H}.

Hence, it suffices to estimate \mathbb{E}[\tilde{B}^{H, v, i}_s \tilde{B}^{H, v, i}_{s, t}]. Recall Y^{H, v} from (?). Since Y^{H, v} and \tilde{B}^{H, v} are independent and B^H = Y^{H, v} + \tilde{B}^{H, v}, we have

As for B^H, the computation is explicit, and it is easy to see that the left-hand side is bounded by (t-s)^{2H} up to constant.

Therefore, it suffices to prove \abs{\mathbb{E}[Y_s^{H, v, i} Y_{s, t}^{H, v, i}]} \lesssim (s-v)^{2H - 1} (t-s). However, this follows from the bound

which in turn follows from Lem. B.4.

Proof of 2. We have

Since

it suffices to prove \abs{\mathbb{E}[(Y^{H, v, i}_t)^2] - \mathbb{E}[(Y^{H, v, i}_s)^2]} \lesssim (s-v)^{2H - 1} (t-s). However, this follows from the bound

which follows from Lem. B.4.

We essentially repeat the argument of [matsuda22, Prop. 3.5]. For simplicity, we drop scripts on H. We first consider the estimate (?). We compute

and

In view of (?), it is easy to see the estimate (?).

We turn to the estimate (?). We recall Y^{v} from (?) and Y^{\tau, v} from (Proof), and we recall the decomposition B = Y^v + \tilde{B}^v and \tilde{B}^{\tau} = Y^{\tau, v} + \tilde{B}^v, where Y^v and Y^{\tau, v} are measurable with respect to \mathcal{F}_v and \tilde{B}^v is independent of \mathcal{F}_v. We set

with \rho = \rho_H defined by (?). As in [matsuda22, Prop. 3.5], especially [matsuda22, Eq. (3.8) and (3.11)], we have

\mathbb{E}[\delta A_{s, u, t} \vert \mathcal{F}_v] = D^0_{s, u, t} + D^i_{s, u, t},

where

The estimate as in [matsuda22, Eq. (3.9)] gives

To estimate \rho(v, t) - \rho(v, s), we observe that

where the last inequality follows from Lem. B.5. Using the bound on \rho from (?) and that on Y^{\tau, v} from (?), we obtain

Since

(s-v)^{(\gamma + 1)H -2} (t-s)^2 \leq (s-v)^{(\gamma - 1)H - 1} (t-s)^{1 + 2H},

we have

To estimate D^i_{s, u, t}, as in [matsuda22] (see the two estimates before (3.12) therein), we have

Therefore, estimating Y^v and \rho(v, t) - \rho(v, u) as before, we obtain

Combining the estimate of \mathbb{E}[\tilde{B}^{v, i}_u \tilde{B}^{v, i}_{u, t}] from Lem. B.5-(i), we obtain

Now the estimate (?) is proven.