We prove an existence and uniqueness result for stochastic differential delay equations driven by a fractional Brownian motion with Hurst parameter H>21, where the delay is a random process. Extending the pathwise Riemann–Stieltjes approach of Ferrante and Rovira \cite{FR2010}, we show that if the random delay τ(⋅,ω) has almost surely γ-Hölder continuous trajectories with γδ>(1−H)/H, where δ∈(0,1] is the Hölder exponent of the gradient of the diffusion coefficient, then the equation has a unique solution with paths in C1−α(−r0,T;Rd)⊂W0α,∞(−r0,T;Rd) for appropriate α∈(1−H,21). We also prove that the solution has moments of all orders when γ>1/(2H) and the Hölder constant of the delay is bounded.
Introduction
Stochastic differential delay equations (SDDEs) arise naturally in mathematical finance, population dynamics, control theory, and neuroscience, wherever the present state of a system depends on its past.
In [4], Ferrante and Rovira proved existence, uniqueness and moment estimates for the equation
with a fixed delay r>0, using the Riemann–Stieltjes integral in the sense of Zähle [20]. In [5] they studied the convergence of solutions of (1) as r→0.
But it is well-known that sometimes delays are not constants, for instance in neural networks [18], in population biology and epidemiology [1, 13], in mathematical finance [2], or in dynamical systems [6].
Our goal is to extend the results of [5] to the equation
where τ:[0,T]×Ω→[0,r0] is a random delay with almost surely Hölder-continuous trajectories, r0>0 is the maximum delay, and η:[−r0,0]→Rd is the initial path. Here both coefficients b and σ depend pointwise on the current value X(s) and the delayed value X(s−τ(s)).
When the driving noise is a standard Brownian motion, the theory on delays, based on Itô calculus, is by now well developed. Stochastic functional differential equations with general path-dependent coefficients are treated in the monographs of Mohammed [14, 15] and Mao [11]. Random or state-dependent delays in the Itô setting have also attracted attention. Random delays appear in Mohammed [14] and in Kolmanovskii and Myshkis [8]. Mao, Shen and Yuan [12] study the almost sure asymptotic stability of neutral stochastic differential delay equations with Markovian switching.
For the fractional Brownian case, the literature on delays is much thinner. Ferrante and Rovira [4, 5] study delay equations by classical methods, using a pathwise Riemann–Stieltjes approach. Neuenkirch, Nourdin and Tindel [16] treat delay equations driven by rough paths, which cover fBm with H>31, using the Lyons rough-path theory [10], and León and Tindel [9] study Malliavin calculus for fractional delay equations with fixed delay. As far as the author knows, it is the first study on random delays in the fBm setting.
In this paper, since H>21, the integral ∫0tσdWH is defined pathwise as a Riemann–Stieltjes integral; see Young [19], Zähle [20] and Nualart–Răşcanu [17]. For each fixed ω, WH(⋅,ω) is a (H−ε)-Hölder function, while τ(⋅,ω) is assumed to be γ-Hölder with γδ>(1−H)/H. Under these assumptions we work pathwise: for each ω we solve a deterministic integral equation with the specific realizations of (WH(⋅,ω),τ(⋅,ω)). As in [17, 5], the stochastic result is then obtained by applying the deterministic one for almost every ω.
The main difference with respect to [17, 5] is that we work with Hölder seminorms, following [7], and we use a contraction argument only on short time intervals. In the above papers the solution is obtained as the fixed point of a contraction on the whole interval [0,T], with respect to the weighted norms of the spaces W0α,∞, but when the delay is variable, we do not know how to control the delayed term X(s−τ(s)) in these norms.
The paper is organized as follows. Section 2 introduces the function spaces, defines the Riemann–Stieltjes integral, and states the hypotheses on the coefficients. Section 3 states the main result, Theorem 3.1. Section 4 contains the key new Lemma 4.1, which controls the regularity of the composition s↦f(s−τ(s)). Section 5 proves existence and uniqueness for the deterministic version of (2), together with an explicit a priori bound. Section 6 applies these results to the stochastic equation and completes the proof of Theorem 3.1.
Preliminaries
Function spaces
Fix parameters r0>0, T>0 and α∈(0,21). We first recall the Hölder spaces.
Definition 2.1. For λ∈(0,1] and an interval I⊂R, the Hölder space Cλ(I;Rd) consists of all continuous functions f:I→Rd such that the seminorm
[f]Cλ(I):=s,t∈Is=tsup∣t−s∣λ∣f(t)−f(s)∣<∞.
We write Cλ for Cλ([−r0,T];Rd) when the interval and target are clear, and ∥f∥Cλ(I):=∥f∥L∞(I)+[f]Cλ(I).
For functions defined only on [0,T], we write Wα,∞(0,T;Rd) with the same formula but integration from 0, and we denote the corresponding norm by ∥f∥α,0.
Remark 2.3. Notice that if f is β-Hölder continuous with β>α, then
In particular, C1−α([−r0,T];Rd)⊂W0α,∞(−r0,T;Rd) with
∥f∥α,r0≤Cα,T,r0∥f∥C1−α([−r0,T]).(4)
The Riemann–Stieltjes integral
The Riemann–Stieltjes integral ∫0tf(s)dgs is not well-defined for all pairs (f,g) of Hölder functions. The theory of Nualart and Răşcanu [17], following Zähle [20], provides sufficient conditions via fractional calculus.
For the definition of the integral ∫fdg and of the constant Λα(g) associated with a function g:[0,T]→Rm and α∈(0,21) we refer to [17]; here we only recall the two facts that will be used.
Remark 2.4. If g∈Cβ′([0,T];Rm) with β′>1−α, then, as shown in [17], Λα(g)<∞ and, more precisely,
Λα(g)≤Cα,β′,T[g]Cβ′([0,T]).(5)
In particular, for g=WH with H>21, choosing α∈(1−H,21) ensures Λα(WH)<∞ a.s.
The estimate that we shall use is the following one, from [17 p. 61]. Let g satisfy Λα(g)<∞, let 0≤u<s≤T and let f:[u,s]→Rd×m be measurable with
Moreover the integral is additive with respect to the interval of integration: ∫usfdg=∫uvfdg+∫vsfdg for u<v<s. When g=(g1,…,gm) takes values in Rm and f in Rd×m, the integral is understood as ∫fdg=∑j=1m∫f⋅jdgj, where f⋅j is the j-th column of f, and we set Λα(g):=∑j=1mΛα(gj). Since ∣f⋅j∣≤∣f∣, the estimate (6) then holds as stated, by applying the scalar estimate to each term of the sum.
Hypotheses
For (x,y)∈Rd×Rd we use the norm ∣(x,y)∣:=∣x∣+∣y∣, we write ∇σ:=∇(x,y)σ for the gradient of σ with respect to both spatial arguments, and ∥⋅∥ for the corresponding operator norm.
(H1)σ:[0,T]×Rd×Rd→Rd×m is measurable, σ(t,⋅,⋅) is C1 for each t, and there exist constants β,δ∈(0,1], M0>0 and, for each N≥0, a constant MN>0 such that for all s,t∈[0,T]:
(i) ∣σ(t,x,y)−σ(t,x′,y′)∣≤M0(∣x−x′∣+∣y−y′∣) for all x,y,x′,y′∈Rd;
(ii) ∥∇σ(t,x,y)−∇σ(t,x′,y′)∥≤MN(∣x−x′∣+∣y−y′∣)δ for all x,y,x′,y′ with ∣x∣,∣y∣,∣x′∣,∣y′∣≤N;
(iii) ∣σ(t,x,y)−σ(s,x,y)∣+∥∇σ(t,x,y)−∇σ(s,x,y)∥≤M0∣t−s∣β for all x,y∈Rd.
(H2)b:[0,T]×Rd×Rd→Rd is measurable and there exist a function b0∈L1/α(0,T), a constant L0≥0 and, for each N≥0, a constant LN>0 such that for all t∈[0,T]:
(i) ∣b(t,x,y)−b(t,x′,y′)∣≤LN(∣x−x′∣+∣y−y′∣) for all x,y,x′,y′∈Rd such that ∣x∣,∣y∣,∣x′∣,∣y′∣≤N;
(ii) ∣b(t,x,y)∣≤L0(∣x∣+∣y∣)+b0(t) for all x,y∈Rd.
(Hτ) The random delay τ:[0,T]×Ω→[0,r0] is a measurable stochastic process, and there exist γ∈(0,1] with
γδ>H1−H
and a random variable Cτ≥0, finite almost surely, such that, almost surely,
∣τ(t,ω)−τ(s,ω)∣≤Cτ(ω)∣t−s∣γ,∀s,t∈[0,T].(7)
Remark 2.5. As in [17, 5], the Lipschitz condition (H1)(i) is global. It implies the linear growth bound
Remark 2.6. When δ=1, for instance if σ(t,⋅,⋅) is C2 with locally bounded second derivatives, the condition in (Hτ) reads γ>(1−H)/H. It is satisfied, for example, by any Lipschitz delay, for which γ=1, or by τ(t)=ψ(Yt) with ψ:R→[0,r0] Lipschitz and Y a process with γ-Hölder paths.
Main result
Let (Ω,F,P) be a complete probability space and WH=(WH,j)j=1m independent fractional Brownian motions with Hurst parameter H>21. Recall that WH,j is a centered Gaussian process with covariance E[WH,j(t)WH,j(s)]=21(t2H+s2H−∣t−s∣2H). We consider equation (2) under the hypotheses of § 2.3. The main result of the paper is the following.
Theorem 3.1. Let H>21. Assume (H1) with β>1−H, (H2) and (Hτ), and set
α0:=min{21,β,1+δγδγ}>1−H.
Fix α∈(1−H,α0), assume b0∈L1/α(0,T), and let η be a stochastic process with trajectories in C1−α([−r0,0];Rd) almost surely.
(a) There exists a unique stochastic process X with trajectories in C1−α([−r0,T];Rd) almost surely which is a solution of the stochastic equation (2). In particular,
Remark 3.2. The inequality α0>1−H holds because β>1−H and because δγ/(1+δγ)>1−H is equivalent to δγ>(1−H)/H, which is assumed in (Hτ). For a Lipschitz delay, that is, for γ=1, the first two conditions in (9) hold automatically.
The proof of Theorem 3.1 is given in § 6. It relies on the composition lemma of § 4 and on the deterministic results of § 5.
The composition lemma
The following lemma is the key new ingredient of this paper.
Lemma 4.1. Let α∈(0,21), γ>α/(1−α), r0,T>0. Suppose f∈C1−α(−r0,T;Rd) and τ:[0,T]→[0,r0] is γ-Hölder with constant [τ]γ:=[τ]Cγ([0,T]). Define fτ:[0,T]→Rd by fτ(s):=f(s−τ(s)). Then fτ∈W0α,∞(0,T;Rd); more precisely:
The exponent satisfies γ(1−α)−α−1>−1 if and only if γ(1−α)>α, i.e. γ>α/(1−α), which holds by assumption. The integral equals tγ(1−α)−α/(γ(1−α)−α)≤Tγ(1−α)−α/(γ(1−α)−α), giving (b).
Finally, combining the uniform bound with (b) we obtain (c). □
The deterministic equation
In this section we fix ω∈Ω and work deterministically with the realizations g:=WH(⋅,ω) and τ:=τ(⋅,ω). Thus g:[0,T]→Rm is a fixed function with Λα(g)<∞ and τ:[0,T]→[0,r0] is a fixed γ-Hölder function.
where K0 is the constant in (8). By (14) we have β>α and δγμ>α, so κ>α and c0<∞; note also that every exponent θ∈{1,β,μ,γμ,δγμ} satisfies 1/(θ−α)≤c0.
For an interval I we write ∥f∥∞,I:=supI∣f∣ and [f]μ,I:=[f]Cμ(I). For 0<b≤T and f:[−r0,b]→Rd we set fτ(s):=f(s−τ(s)), Φf(s):=(f(s),fτ(s)), s∈[0,b].
By Lemma 4.1(a) and ∣s−u∣μ≤T(1−γ)μ∣s−u∣γμ≤Tμ∣s−u∣γμ, for all u,s∈[0,b]:
Proof. (a) follows from (H1)(i) and (H1)(iii). Note that (H1)(i) gives ∥∇σ(t,z)∥≤M0 for all z.
(b) Apply the mean value theorem in integral form to the C1 map ψ(z):=σ(t,z)−σ(s,z) on the segment [z3,z1]:
ψ(z1)−ψ(z3)=∫01∇ψ(z3+θ(z1−z3))dθ(z1−z3),
and by (H1)(iii) the gradient of ψ is bounded in norm by M0∣t−s∣β. Here the gradient is taken with respect to both arguments x and y, which is why (H1)(iii) is stated for the full gradient.
(c) By the mean value theorem in integral form, applied in the variable z=(x,y),
Then, by (27), (30) and (31), v satisfies the assumption of Lemma 5.1 on every [u,s]⊂[a,b] with the exponents β, μ, γμ, δγμ, all larger than α by (14). Using (18), (26), the definition of c0 and s−u≤ℓ≤1,
where η∈Cμ([−r0,0];Rd) is the initial path. For T′∈(0,T], a solution on [0,T′] is a function x∈Cμ([−r0,T′];Rd) satisfying (32) for t∈(0,T′]. Clearly, a solution on [0,T′] satisfies, for every 0≤a<b≤T′,
x(t)=x(a)+Fa(x)(t)+Ga(x)(t),t∈[a,b].(33)
The following proposition gives an a priori bound for the solutions of (32).
which are finite because x∈Cμ([−r0,T′]). Let 0≤a<b≤T′ with ℓ′:=b−a≤ℓ and put h:=[x]μ,[a,b]. Since x(t)−x(a) is controlled by h on [a,b], and the Hölder seminorm on the union of two adjacent intervals is at most the sum of the two seminorms,
Ab≤Aa+ℓμh,Hb≤Ha+h,soZb≤Za+2ℓμh.(37)
By (33), h≤[Fa(x)]μ,[a,b]+[Ga(x)]μ,[a,b]. We apply Proposition 5.3 on [a,b], whose length is ℓ′≤ℓ≤1, with f=x, so that ∥f∥∞,[−r0,b]=Ab and [f]μ,[−r0,b]=Hb, and we bound ℓ′ by ℓ:
The first one follows from ℓ≤1/(16L0) and the second one from the third entry of the minimum. Moreover, since 0<ℓ≤1 and γ≤1 we have γμ≤μ and hence
ℓμ≤ℓγμ.
Together with 1+Tμ≥1, the second condition in (39) therefore implies
c0ΛM0ℓμ≤c0ΛM0(1+Tμ)ℓγμ≤161.(40)
Notice that the terms (III) and (IV) carry no power of ℓ in (38). Multiplying (38) by ℓμ and treating each term separately,
ℓμ⋅(I)=2L0ℓμ+αAb=2L0ℓAb≤81Abby μ+α=1 and (39),
ℓμ⋅(II)≤∥b0∥L1/αby ℓ≤1,
ℓμ⋅(III)=c0Λℓμ(K0+M0)≤16M0K0+M0by (40),
ℓμ⋅(IV)=2c0ΛM0ℓμAb≤81Abby (40),
ℓμ⋅(V)≤161ℓμHbby (39).
Adding up, and recalling the definitions of C1 and Zb=Ab+ℓμHb,
ℓμh≤81Ab+81Ab+161ℓμHb+C1≤41Zb+C1.
Inserting this into (37) gives Zb≤Za+21Zb+2C1, that is
Zb+4C1≤2(Za+4C1).(41)
Now split [0,T′] into at most n intervals of length at most ℓ and iterate (41): ZT′+4C1≤2n(Z0+4C1). Since Z0=∥η∥∞+ℓμ[η]μ≤∥η∥Cμ([−r0,0]), this is (35).
Finally, from (34),
ℓ1≤1+16L0+(16c0M0(1+Tμ)Λ)1/(γμ),n≤1+ℓT.(42)
Since μ=1−α and ∥x∥∞,[−r0,T′]+[x]μ,[−r0,T′]≤(1+ℓ−μ)(∥x∥∞,[−r0,T′]+ℓμ[x]μ,[−r0,T′]), the bound (35) and n≤1+T/ℓ give (36). □
Existence and uniqueness: deterministic case
We can now state and prove the existence and uniqueness result for the deterministic equation (32).
Theorem 5.6. Assume (H1) and (H2), let α∈(0,α0) with α0 as in (14), let τ:[0,T]→[0,r0] be γ-Hölder, let g:[0,T]→Rm satisfy Λα(g)<∞ and let η∈C1−α([−r0,0];Rd). Then equation (32) has a unique solution
x∈C1−α([−r0,T];Rd)⊂W0α,∞(−r0,T;Rd),
and x satisfies the bound (36) with T′=T.
Proof. By Proposition 5.5, every solution on any [0,T′], T′≤T, satisfies
∥x∥∞,[−r0,T′]≤N∗,[x]μ,[−r0,T′]≤R∗,
with N∗:=2n(∥η∥Cμ+4C1) and R∗:=ℓ−μN∗, which do not depend on T′ or on the solution. Define
We first prove local existence. Let 0≤a<b≤T with b−a≤ℓ∗ and suppose that x is a solution on [0,a]; if a=0, we take x=η. Let
X:={y∈Cμ([−r0,b];Rd):y=x on [−r0,a],[y]μ,[a,b]≤R},d(y,y′)=[y−y′]μ,[a,b].
The set X is non-empty, since it contains the extension of x by the constant x(a) on [a,b]. Then it is easy to check that (X,d) is a complete metric space, with underlying Banach space Cμ([−r0,b];Rd) with the norm ∥y∥Cμ([−r0,b])=∥y∥∞,[−r0,b]+[y]μ,[−r0,b]. Notice also that every y∈X satisfies, by the a priori bounds above,
the values of the delayed term yτ on [a,b] are values of y on [−r0,b], so ∥yτ∥∞,[a,b]≤N∗+ℓ′μR, and by (15)∣yτ(s)−yτ(u)∣≤Tμ(R∗+R)∣s−u∣γμ for u,s∈[a,b].
Inserting the three bounds above into (20) and (21) with f=y gives
On the other hand, for ER we use (39) and (40), again with ℓ′≤ℓ in place of ℓ, term by term:
2L0ℓ′α+μR=2L0ℓ′R≤81R,
2c0ΛM0ℓ′μR≤81R,
c0ΛM0(1+Tμ)ℓ′γμR≤161R,
so that ER≤(81+81+161)R=165R. Therefore
[L(y)]μ,[a,b]≤E0+ER≤21R+165R=1613R≤R.
Hence L(X)⊂X. By Proposition 5.4, which is applicable with the constants N and R thanks to (44), and by (43),
d(L(y),L(y′))≤Θd(y,y′)≤21d(y,y′),y,y′∈X.
By the Banach fixed point theorem L has a fixed point y∈X. By (33) for x at time a and the additivity of the integrals, y is a solution on [0,b].
Once local existence is proved, global existence follows by an iterative procedure. Starting from a=0 and applying the local existence result successively on the intervals [kℓ∗,(k+1)ℓ∗∧T] we obtain, after finitely many steps, a solution on [0,T]. Note that ℓ∗ is the same at every step, because the constants are uniform.
We finish by proving the uniqueness of the solution. Let x,x′ be two solutions on [0,T] and let t0:=sup{t∈[0,T]:x=x′ on [−r0,t]}. By continuity, x=x′ on [−r0,t0]. Suppose t0<T. Both solutions satisfy the a priori bounds above, so by (33) and Proposition 5.4 with a=t0 and b=(t0+ℓ∗)∧T,
Hence x−x′ is constant on [a,b], and it vanishes at a; so x=x′ on [−r0,b] with b>t0, a contradiction. Therefore x=x′ on [−r0,T].
Proof of the main theorem
Let WH be the fractional Brownian motion of § 3. We know that Λα(WH)<∞ a.s. for all α∈(1−H,21), by Remark 2.4. For α∈(1−H,21) we write Λα:=Λα(WH).
The following integrability result is a consequence of Fernique’s theorem [3].
Proof of Theorem 3.1. (a) Let Ω∗ be the set of full probability on which (7) holds with Cτ(ω)<∞, Λα(WH(⋅,ω))<∞ and η(⋅,ω)∈C1−α([−r0,0];Rd). For every ω∈Ω∗ the assumptions of Theorem 5.6 are satisfied with g=WH(⋅,ω), τ=τ(⋅,ω) and initial path η(⋅,ω). Hence, for every ω∈Ω∗, the deterministic equation (32) has a unique solution
X(⋅,ω)∈C1−α([−r0,T];Rd)⊂W0α,∞(−r0,T;Rd),
and this is precisely equation (2) for that ω. This gives the existence of the solution and its uniqueness among the processes with trajectories in C1−α([−r0,T];Rd).
(b) Let cˉ:=∥Cτ∥L∞(Ω), so that a.s. T(ω)=T1−γ+[τ(⋅,ω)]γ≤T1−γ+cˉ=:Tˉ, a deterministic constant. Write μ=1−α. By (36), and using ℓ≤1, 1+ℓ−μ≤1+ℓ−1≤e1/ℓ and 2T/ℓ≤eT/ℓ,
[1]B. Boukanjime, T. Caraballo, M. El Fatini and M. El Khalifi, Dynamics of a stochastic coronavirus (COVID-19) epidemic model with Markovian switching, Chaos Solitons Fractals 141 (2020), 110361.DOI
[2]Á. Cartea and L. Sánchez-Betancourt, Optimal execution with stochastic delay, Finance Stoch. 27 (2023), 1–47.DOI
[3]X. Fernique, Regularité des trajectoires des fonctions aléatoires gaussiennes, in École d’Été de Probabilités de Saint-Flour, IV-1974, Lecture Notes in Math. 480 (1975), 1–96.DOI
[4]M. Ferrante and C. Rovira, Stochastic delay differential equations driven by fractional Brownian motion with Hurst parameter H > 1/2, Bernoulli 12 (2006), 85–100.
[5]M. Ferrante and C. Rovira, Convergence of delay differential equations driven by fractional Brownian motion, J. Evol. Equ. 10 (2010), 761–783.arxiv.org/abs/0903.5498
[6]M.J. Garrido-Atienza, A. Ogrowsky and B. Schmalfuß, Random differential equations with random delays, Stoch. Dyn. 11 (2011), 675–699.
[7]Y. Hu and D. Nualart, Differential equations driven by Hölder continuous functions of order greater than 1/2, in Stochastic Analysis and Applications (Abel Symposium 2005), Springer, 2007, 399–413.DOI
[8]V. Kolmanovskii and A. Myshkis, Introduction to the Theory and Applications of Functional Differential Equations, Springer, 1999.DOI
[9]J. León and S. Tindel, Malliavin calculus for fractional delay equations, J. Theoret. Probab. 25 (2012), 854–889.arxiv.org/abs/0912.2180
[10]T. Lyons, Differential equations driven by rough signals (I): An extension of an inequality of L.C. Young, Math. Res. Lett. 1 (1994), 451–464.DOI
[11]X. Mao, Stochastic Differential Equations and Applications, Horwood, 1997.
[12]X. Mao, Y. Shen and C. Yuan, Almost surely asymptotic stability of neutral stochastic differential delay equations with Markovian switching, Stochastic Process. Appl. 118 (2008), 1385–1406.DOI
[13]M. Martcheva, N. Tuncer and C. N. Ngonghala, Mathematical models of epidemics with infection time, AppliedMath 5 (2025), no. 2, 47.
[14]S.-E.A. Mohammed, Stochastic Functional Differential Equations, Research Notes in Mathematics 99, Pitman, 1984.
[15]S.-E.A. Mohammed, Stochastic differential systems with memory: theory, examples and applications, in Stochastic Analysis and Related Topics VI, Progr. Probab. 42, Birkhäuser, 1998, 1–77.DOI
[16]A. Neuenkirch, I. Nourdin and S. Tindel, Delay equations driven by rough paths, Electron. J. Probab. 13 (2008), 2031–2068.arxiv.org/abs/0711.2633
[17]D. Nualart and A. Răşcanu, Differential equations driven by fractional Brownian motion, Collect. Math. 53 (2002), 55–81.
[18]J. Touboul, Limits and dynamics of stochastic neuronal networks with random heterogeneous delays, J. Stat. Phys. 149 (2012), 569–597.DOI
[19]L.C. Young, An inequality of the Hölder type connected with Stieltjes integration, Acta Math. 67 (1936), 251–282.DOI
[20]M. Zähle, Integration with respect to fractal functions and stochastic calculus I, Prob. Theory Relat. Fields 111 (1998), 333–374.DOI