Introduction
Backward stochastic differential equations (BSDEs) have become an important tool in stochastic analysis, nonlinear expectations, stochastic control, and mathematical finance. Since the seminal work of Pardoux and Peng [26 ] , the theory has developed well beyond the classical Brownian framework, encompassing more general filtrations, discontinuous martingales, random terminal horizons, and nonlinear constraints.
Reflected backward stochastic differential equations (RBSDEs) constitute a particularly significant extension. Their classical formulation, developed by El Karoui, Kapoudjian, Pardoux, Peng and Quenez [13 ] , provides a probabilistic approach to optimal stopping and obstacle problems. In a Brownian filtration, a reflected solution ( Y , Z , K ) (Y, Z, K) ( Y , Z , K ) typically satisfies
Y t = ξ + ∫ t T f ( s , Y s , Z s ) d s + K T − K t − ∫ t T Z s d W s , Y t ≥ L t , Y_t = \xi+ \int_t^T f(s,Y_s,Z_s)\,\mathrm{d}s + K_T - K_t - \int_t^T Z_s\,\mathrm{d}W_s,\qquad Y_t \ge L_t, Y t = ξ + ∫ t T f ( s , Y s , Z s ) d s + K T − K t − ∫ t T Z s d W s , Y t ≥ L t ,
together with the Skorokhod minimality condition
∫ ( 0 , T ] ( Y t − L t ) d K t = 0. \int_{(0,T]} (Y_t-L_t)\,\mathrm{d}K_t = 0. ∫ ( 0 , T ] ( Y t − L t ) d K t = 0.
The increasing process K K K enforces the obstacle constraint, while the minimality condition ensures that reflection acts only when the solution reaches the obstacle. This contact property is central to the connection between RBSDEs, Snell envelopes, and nonlinear optimal stopping.
Irregular obstacles and generalized reflection. The classical theory commonly relies on regularity assumptions on the obstacle and the solution. Such assumptions can be restrictive in applications involving discontinuous information, irregular exercise rewards, or financial contracts subject to default and early termination.
The work of Grigorova, Imkeller, Offen, Ouknine and Quenez [21 ] develops reflected BSDEs with obstacles that are not necessarily right-continuous. Their approach relies on strong optional supermartingales, Mertens decomposition, and the Gal’chouk–Lenglart formula, and establishes connections with optimal stopping under nonlinear expectations.
Further contributions include the general-filtration analysis of Baadi and Ouknine [5 ] , the predictable-obstacle framework of Bouhadou and Ouknine [7 ] , and the study of doubly reflected BSDEs in the predictable setting by Arharas, Bouhadou and Ouknine [3 ] . Akdim, Haddadi and Ouknine [1 ] also investigate strong Snell envelopes and reflected BSDEs with regulated trajectories.
These developments highlight the importance of distinguishing optional and predictable reflection. For irregular optional obstacles, the reflection mechanism may involve a predictable increasing process and a purely discontinuous optional increasing process, whose minimality conditions are formulated using different contact conventions. Consequently, the classical continuous-obstacle Skorokhod condition cannot simply be transferred to arbitrary optional or predictable obstacles.
Default times and reflected BSDEs
A complementary extension of BSDE theory arises when the model includes a random time τ \tau τ , representing, for example, the default of a counterparty, the failure of a system, or the termination of a financial contract.
In credit-risk models, the reference filtration F \mathbb{F} F describes the information available before default, while an enlarged filtration G \mathbb{G} G incorporates the observation of τ \tau τ . The default time is then a G \mathbb{G} G -stopping time, although it need not be an F \mathbb{F} F -stopping time.
The mathematical foundations of default modeling through enlargement of filtrations were developed in particular by Bielecki, Jeanblanc and Rutkowski [6 ] , and Jeanblanc and Le Cam [22 ] . The conditional-density approach of El Karoui, Jeanblanc and Jiao [14 ] and the survival-process approach of Jeanblanc and Song [25 ] provide further tools for studying random times and their associated semimartingale decompositions.
The martingale representation problem in enlarged filtrations is particularly important for backward stochastic equations. Jeanblanc and Song [24 ] investigate the martingale representation property under progressive enlargement, while Aksamit, Jeanblanc and Rutkowski [2 ] develop integral representation results for martingales stopped at a random time. Related enlargement formulas for multiple, possibly non-ordered default times are established by Jeanblanc, Li and Song [23 ] .
In this setting, a BSDE may be driven by both a Brownian motion and a compensated default martingale. The latter introduces a jump coefficient and a compensator that must be taken into account when comparing equations in the reference and enlarged filtrations.
The analysis of BSDEs with default jumps and their applications to nonlinear pricing and optimal stopping has been developed, among others, by Dumitrescu, Grigorova, Quenez and Sulem [12 ] , and by Dumitrescu, Quenez and Sulem [11 ] (see also [4 ] ).
More recently, Elmansouri [15 ] established existence and uniqueness for reflected BSDEs with a default time and an obstacle that is not necessarily right-continuous, under stochastic Lipschitz assumptions. In a related study, Elmansouri [16 ] examined BSDEs and RBSDEs with default time and stochastic Lipschitz drivers, together with applications to the pricing of European and American options.
Elmansouri and El Otmani [18 ] investigated optimal control over split stopping times in a defaultable framework, establishing connections with reflected BSDEs with irregular obstacles and nonlinear expectations. Their work [17 ] also extends reflected BSDEs with default times to settings involving Brownian and jump noise, with applications to optimal stopping and the valuation of American options.
In addition, Elmansouri and El Otmani [19 ] studied well-posedness for default-time RBSDEs with irregular barriers and their relationship with nonlinear optimal stopping. These contributions provide existence, uniqueness, and control-theoretic results for specific defaultable RBSDE formulations.
Random horizons, optimal stopping, and filtration reduction
Beyond the existence theory of BSDEs and RBSDEs, a separate line of research studies how optimal stopping and valuation problems can be transferred between the reference filtration and its progressive enlargement.
Choulli and Alsheyab [8 ] establish relationships between optimal stopping problems under a random horizon and corresponding reference-filtration problems. Their results include the parametrization of enlarged-filtration reward processes, representations of Snell envelopes, and connections between optimal stopping times.
Related questions for two-player stopping problems are addressed by Elmansouri, Marzougue and El Jamali [20 ] , who examine Dynkin games under progressive enlargement and establish corresponding reference-filtration reduction results.
A further important framework is the theory of invariance times developed by Crépey and Song [10 ] . Under suitable filtration and measure-change hypotheses, this theory provides a mechanism for transporting reference-filtration local martingales, stopped immediately before a random time, into the enlarged filtration.
However, martingale transport and reflection minimality are different mathematical properties. Even when the martingale component of a BSDE admits a suitable transformation, this does not automatically establish that its finite-variation reflection measure preserves the required contact condition.
The reflection-support problem
The additional difficulty can be described directly in terms of the random measure associated with an increasing reflection process.
Consider an F \mathbb{F} F -predictable increasing process A A A and a nonnegative predictable contact integrand H H H . Suppose that the enlarged-filtration formulation provides
∫ ( 0 , T ] H t 1 { t ≤ τ } d A t = 0 a.s. \int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,dA_t=0 \qquad\text{a.s.} ∫ ( 0 , T ] H t 1 { t ≤ τ } d A t = 0 a.s.
To recover the reference-filtration minimality condition, one needs to deduce
∫ ( 0 , T ] H t d A t = 0 a.s. \int_{(0,T]} H_t\,dA_t=0 \qquad\text{a.s.} ∫ ( 0 , T ] H t d A t = 0 a.s.
This implication is not automatic: stopping at τ \tau τ may remove part of the support of the reflection measure.
Writing q q q for the predictable survival projection, the corresponding projection identity takes the form
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d A t = E ∫ ( 0 , T ] H t q t d A t . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,dA_t
=
\mathbb{E}\int_{(0,T]} H_tq_t\,dA_t. E ∫ ( 0 , T ] H t 1 { t ≤ τ } d A t = E ∫ ( 0 , T ] H t q t d A t .
The relevant positivity criterion is therefore expressed relative to the mean measure induced by d A dA d A , rather than merely relative to Lebesgue measure. This distinction becomes essential when the increasing process contains atoms or singular components.
For optional contact conditions, the corresponding analysis involves the dual optional projection of the stopped reflection measure and the strict pre-default interval. The difference between the indicators 1 { t < τ } \mathbf{1}_{\{t<\tau\}} 1 { t < τ } and 1 { t ≤ τ } \mathbf{1}_{\{t\leq\tau\}} 1 { t ≤ τ } must also be taken into account whenever the reflection measure may charge the graph of τ \tau τ .
These observations motivate a systematic study of the support properties of reflection measures under random horizons.
Main contributions of the paper
The present paper focuses on the measure-theoretic conditions ensuring that reflection minimality is preserved under stopping, and on their interaction with the reconstruction of reflected BSDEs. Our contributions can be summarized as follows.
First, we establish sharp optional and predictable contact-transfer criteria using dual projections and the mean measures associated with increasing processes. The resulting conditions identify when stopped and unstopped contact identities are equivalent, including when the reflection measures possess atomic or singular components.
Second, we specialize the analysis to Cox random times. The explicit conditional survival structure yields strong contact equivalence, and we extend the survival-weighting principle to general random measures determined by the reference information. Under uniform survival positivity, we also obtain quantitative estimates controlling approximate contact errors.
Third, in a Brownian Cox framework, we establish an explicit reconstruction of a reflected BSDE with continuous obstacle from a reference-filtration reflected equation. The transformed generator incorporates the default intensity and prescribed recovery, while the enlarged solution is reconstructed through its continuous martingale, default-jump, and reflection components. The converse is obtained within an explicitly specified reduction class, without claiming uniqueness among arbitrary enlarged-filtration solutions.
Fourth, for irregular obstacles, we apply the contact-transfer criteria separately to the predictable and purely discontinuous optional reflection components. We clarify the relevant endpoint and pointwise jump-contact conventions, while distinguishing these minimality statements from a complete Mertens reconstruction theorem.
The novelty of the present analysis thus lies in isolating the exact support conditions required for reflection-contact transfer and combining them with a concrete Cox reconstruction mechanism. It does not consist in asserting general existence and uniqueness for default-time RBSDEs with irregular obstacles, which have already been investigated in the aforementioned literature.
Scope and organization
Our results do not establish an unconditional reduction–reconstruction theorem for arbitrary optional obstacles or general non-Cox invariance times. Such an extension requires additional hypotheses concerning martingale representation, compensator transport, obstacle reduction, default jumps, and the reconstruction of Mertens reflection components.
Section 2 recalls the probabilistic and measure-theoretic framework. Section 3 establishes the optional and predictable contact-transfer criteria, and Section 4 develops their Cox specialization, including general reference random measures.
The Brownian Cox reconstruction theorem is presented in Section 5. The subsequent sections investigate irregular Mertens contact conditions, quantitative stability, and the remaining questions concerning invariance-time reconstruction beyond the Cox setting.
Measure-theoretic preliminaries
This section introduces the probabilistic framework and the measure-theoretic tools used throughout the paper. Particular attention is devoted to filtration enlargement, random Stieltjes measures, and dual optional and predictable projections. These notions will be essential for studying the transfer of reflection contact conditions across random horizons.
Probabilistic framework and filtration enlargement
Let ( Ω , A , P ) (\Omega, \mathcal{A}, \mathbb{P}) ( Ω , A , P ) be a complete probability space endowed with a reference filtration
F = ( F t ) t ≥ 0 \mathbb{F} = (\mathcal{F}_t)_{t\geq0} F = ( F t ) t ≥ 0
satisfying the usual conditions.
We consider a second filtration
G = ( G t ) t ≥ 0 , \mathbb{G} = (\mathcal{G}_t)_{t\geq0}, G = ( G t ) t ≥ 0 ,
also satisfying the usual conditions, such that
F t ⊆ G t , t ≥ 0. (1) \mathcal{F}_t \subseteq\mathcal{G}_t, \qquad t \geq0. \tag*{(1)} F t ⊆ G t , t ≥ 0. (1)
Let τ \tau τ be a nonnegative G \mathbb{G} G -stopping time, not necessarily an F \mathbb{F} F -stopping time, and fix a finite deterministic horizon T > 0 T > 0 T > 0 .
The inclusion (1) describes a situation in which the enlarged filtration G \mathbb{G} G contains additional information relative to the reference filtration F \mathbb{F} F .
Two classical enlargement mechanisms are particularly relevant.
Progressive enlargement. Let τ \tau τ be a nonnegative random time, not necessarily an F \mathbb{F} F -stopping time.
The progressive enlargement of F \mathbb{F} F by τ \tau τ is the smallest right-continuous filtration containing F \mathbb{F} F for which τ \tau τ becomes a stopping time.
It can be defined by
G t = ⋂ s > t ( F s ∨ σ ( τ ∧ s ) ) , t ≥ 0 , \mathcal{G}_t = \bigcap_{s>t} \left(\mathcal{F}_s \vee\sigma(\tau\wedge s)\right), \qquad t \geq0, G t = s > t ⋂ ( F s ∨ σ ( τ ∧ s ) ) , t ≥ 0 ,
with the usual completion.
In particular,
{ τ ≤ t } ∈ G t , t ≥ 0 , \{\tau\leq t\} \in\mathcal{G}_t, \qquad t \geq0, { τ ≤ t } ∈ G t , t ≥ 0 ,
whereas this event need not belong to F t \mathcal{F}_t F t .
Thus, the additional information concerns the occurrence of the random time and is revealed progressively.
Initial enlargement. Let L L L be an A \mathcal{A} A -measurable random variable representing additional information available at time zero.
The initial enlargement of F \mathbb{F} F by L L L is defined by
G t = ⋂ s > t ( F s ∨ σ ( L ) ) , t ≥ 0 , \mathcal{G}_t = \bigcap_{s>t} \left(\mathcal{F}_s \vee\sigma(L)\right), \qquad t \geq0, G t = s > t ⋂ ( F s ∨ σ ( L ) ) , t ≥ 0 ,
again with the usual completion.
In this construction, the entire random variable L L L is available from the beginning, in contrast to progressive enlargement, where information about τ \tau τ is revealed over time.
The two mechanisms may lead to different semimartingale decompositions and martingale representation properties.
For instance, an F \mathbb{F} F -Brownian motion need not remain a Brownian motion in an initially or progressively enlarged filtration.
Its decomposition may involve an additional finite-variation term, often referred to as an information drift, whenever the corresponding semimartingale property holds.
Random Stieltjes measures
Throughout the paper, unless otherwise stated, an increasing process a = ( a t ) 0 ≤ t ≤ T a=(a_t)_{0\leq t\leq T} a = ( a t ) 0 ≤ t ≤ T is assumed to be nonnegative, F \mathbb{F} F -adapted, right-continuous, and such that
a 0 = 0 , E [ a T ] < ∞ . (2) a_0=0,\qquad\mathbb{E}[a_T]<\infty. \tag*{(2)} a 0 = 0 , E [ a T ] < ∞. (2)
Such a process induces a finite positive random Stieltjes measure on ( 0 , T ] (0,T] ( 0 , T ] , defined pathwise by
d a ( ω , ( s , t ] ) = a t ( ω ) − a s ( ω ) , 0 ≤ s < t ≤ T . da(\omega,(s,t])=a_t(\omega)-a_s(\omega),\qquad0\leq s<t\leq T. d a ( ω , ( s , t ]) = a t ( ω ) − a s ( ω ) , 0 ≤ s < t ≤ T .
For every nonnegative measurable process H H H for which the integral is well-defined, we use the notation
H ⋅ d a : = ∫ ( 0 , T ] H t d a t . H\cdot da:=\int_{(0,T]}H_t\,da_t. H ⋅ d a := ∫ ( 0 , T ] H t d a t .
An important feature is that the random measure d a da d a need not be absolutely continuous with respect to Lebesgue measure d t dt d t .
In particular, d a da d a may possess absolutely continuous, atomic, or singular continuous components. The following examples illustrate these different possibilities.
Here α \alpha α is a nonnegative, progressively measurable process satisfying the required integrability conditions, t 0 ∈ ( 0 , T ] t_0\in(0,T] t 0 ∈ ( 0 , T ] is deterministic, δ t 0 \delta_{t_0} δ t 0 denotes the Dirac measure at t 0 t_0 t 0 , and C C C is the classical Cantor function on [ 0 , 1 ] [0,1] [ 0 , 1 ] , with T = 1 T=1 T = 1 .
More generally, for almost every ω \omega ω , the finite Stieltjes measure d a ( ω , d t ) da(\omega,dt) d a ( ω , d t ) admits a Lebesgue decomposition
d a ( ω , d t ) = d a a c ( ω , d t ) + d a s ( ω , d t ) , da(\omega,dt)=da^{ac}(\omega,dt)+da^s(\omega,dt), d a ( ω , d t ) = d a a c ( ω , d t ) + d a s ( ω , d t ) ,
where d a a c da^{ac} d a a c is absolutely continuous with respect to d t dt d t and d a s da^s d a s is singular.
The singular component may itself contain an atomic part and a singular continuous part. This distinction is important in the theory of reflected backward stochastic differential equations. Indeed, the increasing reflection process is not generally absolutely continuous with respect to time. Even when the reflection process is continuous, its associated measure may concentrate on a set of zero Lebesgue measure. When discontinuous or irregular obstacles are considered, reflection may also involve increasing components carrying atomic mass. Consequently, contact and minimality conditions must be formulated with respect to the reflection measure itself rather than only with respect to Lebesgue measure.
Financial interpretation
Filtration enlargement provides a natural framework for describing the impact of additional information on the valuation and management of financial risks.
In credit-risk modeling, the reference filtration F \mathbb{F} F may represent the market information generated by observable economic and financial risk factors.
A random time τ \tau τ may represent the default time of a counterparty or the termination time of a financial contract.
The progressively enlarged filtration G \mathbb{G} G then incorporates the observation of default as it occurs. A defaultable financial claim with contractual maturity T T T may terminate at the random horizon
In contrast, initial enlargement may describe an agent who knows an additional random signal concerning future economic conditions or a terminal market variable from time zero. Such additional information may change the semimartingale decomposition of price processes and, consequently, the dynamics of the associated backward stochastic differential equations.
Reflected BSDEs arise naturally in the valuation of financial contracts involving an early exercise or termination option. The obstacle represents the exercise value or another contractual constraint, while the increasing reflection process ensures that the value process respects that constraint. The reflection measure records the times at which the constraint becomes binding. When the contract terminates at τ \tau τ , only the portion of the reflection measure associated with the relevant pre-default interval contributes to the stopped equation.
Example 2.2 (Defaultable claim with early exercise). Consider a financial contract with maturity T T T and counterparty default time τ \tau τ .
Let ξ \xi ξ denote its promised terminal payoff in the absence of default, and let R τ R_\tau R τ denote the recovery payment when default occurs before or at maturity.
The payoff at the random horizon T ∧ τ T \wedge\tau T ∧ τ can be written as
Ξ = ξ 1 { τ > T } + R τ 1 { τ ≤ T } . \Xi= \xi1_{\{\tau>T\}} + R_\tau1_{\{\tau\le T\}}. Ξ = ξ 1 { τ > T } + R τ 1 { τ ≤ T } . If the contract also includes an early exercise feature, its valuation may be formulated through a reflected BSDE stopped at T ∧ τ T \wedge\tau T ∧ τ .
The obstacle represents the exercise payoff, and the reflection process enforces the corresponding value constraint. The random horizon determines when the contract terminates, whereas the reflection measure determines when intervention is required to maintain the obstacle constraint.
This example motivates the study of contact measures restricted to [ 0 , τ ) [0,\tau) [ 0 , τ ) or [ 0 , τ ] [0,\tau] [ 0 , τ ] .
Dual optional and predictable projections
We next recall the dual optional and predictable projections of integrable random measures relative to the reference filtration F \mathbb{F} F .
Let μ \mu μ be a nonnegative random measure on ( 0 , T ] (0,T] ( 0 , T ] satisfying the standard measurability assumptions for dual projection and
E [ μ ( ( 0 , T ] ) ] < ∞ . \mathbb{E}[\mu((0,T])] < \infty. E [ μ (( 0 , T ])] < ∞.
Its dual optional projection relative to F \mathbb{F} F , denoted by μ o , F \mu^{o,\mathbb{F}} μ o , F , is the F \mathbb{F} F -optional random measure characterized by
E [ ∫ ( 0 , T ] H t μ ( d t ) ] = E [ ∫ ( 0 , T ] H t μ o , F ( d t ) ] . \mathbb{E}\left[\int_{(0,T]} H_t \,\mu(dt)\right]
=
\mathbb{E}\left[\int_{(0,T]} H_t \,\mu^{o,\mathbb{F}}(dt)\right]. E [ ∫ ( 0 , T ] H t μ ( d t ) ] = E [ ∫ ( 0 , T ] H t μ o , F ( d t ) ] .
for every bounded nonnegative F \mathbb{F} F -optional process H H H .
Similarly, the dual predictable projection μ p , F \mu^{p,\mathbb{F}} μ p , F is characterized by
E [ ∫ ( 0 , T ] H t μ ( d t ) ] = E [ ∫ ( 0 , T ] H t μ p , F ( d t ) ] . \mathbb{E}\left[\int_{(0,T]} H_t \,\mu(dt)\right]
=
\mathbb{E}\left[\int_{(0,T]} H_t \,\mu^{p,\mathbb{F}}(dt)\right]. E [ ∫ ( 0 , T ] H t μ ( d t ) ] = E [ ∫ ( 0 , T ] H t μ p , F ( d t ) ] .
for every bounded nonnegative F \mathbb{F} F -predictable process H H H .
Both identities extend to arbitrary nonnegative admissible integrands by monotone approximation.
It is important to distinguish these dual projections of random measures from the ordinary optional and predictable projections of stochastic processes.
In particular, if
A t = μ ( ( 0 , t ] ) , A_t = \mu((0,t]), A t = μ (( 0 , t ]) ,
the dual optional projection of μ \mu μ is not, in general, obtained by taking the pointwise optional projection of the cumulative process A A A .
The dual projection is characterized by the preservation of expected integrals against the corresponding class of optional or predictable test processes.
Mean measures and zero-contact conditions
Let H \mathcal{H} H denote either the F \mathbb{F} F -optional sigma-field O ( F ) \mathcal{O}(\mathbb{F}) O ( F ) or the F \mathbb{F} F -predictable sigma-field P ( F ) \mathcal{P}(\mathbb{F}) P ( F ) , restricted to Ω × ( 0 , T ] \Omega\times(0,T] Ω × ( 0 , T ] .
For an increasing process a a a satisfying (2) , define its mean measure on H \mathcal{H} H by
m a H ( B ) = E [ ∫ ( 0 , T ] 1 B ( ω , t ) d a t ( ω ) ] , B ∈ H . m_a^{\mathcal{H}}(B) = \mathbb{E}\left[\int_{(0,T]} \mathbf{1}_B(\omega,t)\, da_t(\omega)\right], \qquad B \in\mathcal{H}. m a H ( B ) = E [ ∫ ( 0 , T ] 1 B ( ω , t ) d a t ( ω ) ] , B ∈ H .
This is a finite positive measure since
m a H ( Ω × ( 0 , T ] ) = E [ a T ] < ∞ . m_a^{\mathcal{H}}(\Omega\times(0,T]) = \mathbb{E}[a_T] < \infty. m a H ( Ω × ( 0 , T ]) = E [ a T ] < ∞.
For every nonnegative H \mathcal{H} H -measurable process H H H , we have
E [ ∫ ( 0 , T ] H t d a t ] = ∫ Ω × ( 0 , T ] H ( ω , t ) m a H ( d ω , d t ) . \mathbb{E}\left[\int_{(0,T]} H_t\, da_t\right] = \int_{\Omega\times(0,T]} H(\omega,t)\,m_a^{\mathcal{H}}(d\omega,dt). E [ ∫ ( 0 , T ] H t d a t ] = ∫ Ω × ( 0 , T ] H ( ω , t ) m a H ( d ω , d t ) .
Since both sides are nonnegative, we obtain the following equivalence:
∫ ( 0 , T ] H t d a t = 0 a.s. ⟺ H = 0 m a H -a.e. \int_{(0,T]} H_t\, da_t = 0 \ \text{a.s.} \quad\Longleftrightarrow\quad H = 0 \quad m_a^{\mathcal{H}}\text{-a.e.} ∫ ( 0 , T ] H t d a t = 0 a.s. ⟺ H = 0 m a H -a.e.
Thus, a zero-contact condition can be expressed as a null-set property of the mean measure associated with the increasing process.
This formulation will be particularly useful when comparing stopped and unstopped reflection measures.
Lemma 2.3 (Zero-integral principle). Let X X X and Y Y Y be nonnegative random variables, possibly taking values in [ 0 , ∞ ] [0,\infty] [ 0 , ∞ ] , such that
E [ X ] = E [ Y ] . \mathbb{E}[X] = \mathbb{E}[Y]. E [ X ] = E [ Y ] . Then
X = 0 a.s. ⟺ Y = 0 a.s. X = 0 \quad\text{a.s.} \quad\Longleftrightarrow\quad Y = 0 \quad\text{a.s.} X = 0 a.s. ⟺ Y = 0 a.s.
Proof. Since X X X and Y Y Y are nonnegative, each variable vanishes almost surely if and only if its expectation is zero.
Thus,
X = 0 a.s. ⟺ E [ X ] = 0. X = 0 \quad\text{a.s.} \quad\Longleftrightarrow\quad\mathbb{E}[X] = 0. X = 0 a.s. ⟺ E [ X ] = 0.
By the assumed equality of expectations,
E [ X ] = 0 ⟺ E [ Y ] = 0. \mathbb{E}[X] = 0 \quad\Longleftrightarrow\quad\mathbb{E}[Y] = 0. E [ X ] = 0 ⟺ E [ Y ] = 0.
Finally, nonnegativity of Y Y Y gives
E [ Y ] = 0 ⟺ Y = 0 a.s. \mathbb{E}[Y] = 0 \Longleftrightarrow Y = 0 \quad\text{a.s.} E [ Y ] = 0 ⟺ Y = 0 a.s.
Combining these equivalences proves the result.
Example 2.5 (Positivity relative to the reflection measure). Let T = 2 T = 2 T = 2 and consider a t = 1 { t ≥ 1 } a_t = \mathbf{1}_{\{t \ge1\}} a t = 1 { t ≥ 1 } . Its associated Stieltjes measure is d a t = δ 1 ( d t ) da_t = \delta_1(dt) d a t = δ 1 ( d t ) .
Define the nonnegative deterministic process q t = 1 { t ≠ 1 } q_t = \mathbf{1}_{\{t \ne1\}} q t = 1 { t = 1 } . Although q q q is positive Lebesgue-almost everywhere,
∫ ( 0 , 2 ] q t d a t = q 1 = 0 , ∫ ( 0 , 2 ] d a t = 1. \int_{(0,2]} q_t\,da_t = q_1 = 0, \qquad\int_{(0,2]} da_t = 1. ∫ ( 0 , 2 ] q t d a t = q 1 = 0 , ∫ ( 0 , 2 ] d a t = 1. Thus, positivity with respect to Lebesgue measure does not imply positivity with respect to the reflection measure. The relevant condition for transferring contact identities must therefore be formulated relative to the measure induced by d a da d a .
The preceding definitions and observations provide the measure-theoretic foundation for the optional and predictable contact transfer criteria developed in the next section.
Exact transfer of optional and predictable contact
We now investigate the transfer of contact conditions between an increasing process in the reference filtration F \mathbb{F} F and its restriction to the stochastic interval determined by a random time τ \tau τ .
The main difficulty is that a contact condition imposed before τ \tau τ need not determine the behavior of the reflection measure on the entire deterministic horizon [ 0 , T ] [0,T] [ 0 , T ] .
Throughout this section, a a a satisfies the standing assumptions of Section 2.
We introduce the two stopped random measures
μ a < ( d t ) = 1 { t < τ } d a t , μ a ≤ ( d t ) = 1 { t ≤ τ } d a t . \mu_a^{<}(dt) = \mathbf{1}_{\{t<\tau\}}\,da_t, \qquad\mu_a^{\le}(dt) = \mathbf{1}_{\{t\le\tau\}}\,da_t. μ a < ( d t ) = 1 { t < τ } d a t , μ a ≤ ( d t ) = 1 { t ≤ τ } d a t .
These measures differ only at the random time τ \tau τ . More precisely, for every nonnegative measurable process H H H ,
∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = ∫ ( 0 , T ] H t 1 { t < τ } d a t + H τ Δ a τ 1 { 0 < τ ≤ T } , (3) \int_{(0,T]} H_t1_{\{t\leq\tau\}}\,da_t
=
\int_{(0,T]} H_t1_{\{t<\tau\}}\,da_t+H_\tau\Delta a_\tau1_{\{0<\tau\leq T\}}, \tag*{(3)} ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = ∫ ( 0 , T ] H t 1 { t < τ } d a t + H τ Δ a τ 1 { 0 < τ ≤ T } , (3)
where Δ a t = a t − a t − \Delta a_t=a_t-a_{t-} Δ a t = a t − a t − is the left-jump size at t t t .
Thus, in view of (3) , the two endpoint conventions agree whenever a a a is continuous, but they may differ when the reflection measure has an atom at τ \tau τ .
We first examine optional contact conditions and then turn to their predictable counterparts.
Optional contact conditions
Let
G t = P ( τ > t ∣ F t ) , t ≥ 0 , (4) G_t=\mathbb{P}(\tau>t\mid\mathcal{F}_t),\qquad t\geq0, \tag*{(4)} G t = P ( τ > t ∣ F t ) , t ≥ 0 , (4)
denote the Azéma supermartingale associated with τ \tau τ relative to F \mathbb{F} F .
More precisely, G G G is the optional projection of the process 1 { t < τ } 1_{\{t<\tau\}} 1 { t < τ } . We use its usual right-continuous version.
The following theorem characterizes when an optional contact condition restricted to [ 0 , τ ) [0,\tau) [ 0 , τ ) can be transferred to [ 0 , T ] [0,T] [ 0 , T ] .
Theorem 3.1 (Optional contact: exact criterion). Let H ≥ 0 H\geq0 H ≥ 0 be an F \mathbb{F} F -optional process, and let ν = ( μ a < ) o , F \nu=(\mu_a^{<})^{o,\mathbb{F}} ν = ( μ a < ) o , F denote the dual optional projection of the stopped random measure μ a < \mu_a^{<} μ a < . Then
∫ ( 0 , T ] H t μ a < ( d t ) = 0 a.s. ⟺ ∫ ( 0 , T ] H t ν ( d t ) = 0 a.s. (5) \int_{(0,T]}H_t\mu_a^{<}(dt)=0\quad\text{a.s.}
\Longleftrightarrow
\int_{(0,T]}H_t\nu(dt)=0\quad\text{a.s.} \tag*{(5)} ∫ ( 0 , T ] H t μ a < ( d t ) = 0 a.s. ⟺ ∫ ( 0 , T ] H t ν ( d t ) = 0 a.s. (5) Moreover, the dual optional projection satisfies
ν ( d t ) = G t d a t , \nu(dt)=G_t\,da_t, ν ( d t ) = G t d a t , in the sense of optional random measures. Consequently,
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t G t d a t = 0 a.s. (6) \int_{(0,T]}H_t1_{\{t<\tau\}}\,da_t=0\quad\text{a.s.}
\Longleftrightarrow
\int_{(0,T]}H_tG_t\,da_t=0\quad\text{a.s.} \tag*{(6)} ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t G t d a t = 0 a.s. (6) Let m a o m_a^o m a o denote the mean measure associated with d a da d a on the F \mathbb{F} F -optional sigma-field O ( F ) \mathcal{O}(\mathbb{F}) O ( F ) , as introduced in Section 2.
Then the following equivalence holds for every nonnegative F \mathbb{F} F -optional process H H H :
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t d a t = 0 a.s. , \int_{(0,T]}H_t1_{\{t<\tau\}}\,da_t=0\quad\text{a.s.}
\Longleftrightarrow
\int_{(0,T]}H_t\,da_t=0\quad\text{a.s.}, ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t d a t = 0 a.s. , if and only if
G > 0 m a o -almost everywhere . G>0\qquad m_a^o\text{-almost everywhere}. G > 0 m a o -almost everywhere .
Proof. We divide the proof into three steps.
Step 1: Dual optional projection.
By the defining property of the dual optional projection, for every bounded nonnegative F \mathbb{F} F -optional process H H H ,
E ∫ ( 0 , T ] H t μ a < ( d t ) = E ∫ ( 0 , T ] H t ν ( d t ) . \mathbb{E}\int_{(0,T]}H_t\mu_a^{<}(dt)
=
\mathbb{E}\int_{(0,T]}H_t\nu(dt). E ∫ ( 0 , T ] H t μ a < ( d t ) = E ∫ ( 0 , T ] H t ν ( d t ) .
For a general nonnegative optional process H H H , define
H t ( n ) = H t ∧ n , n ≥ 1. H_t^{(n)}=H_t\wedge n,\qquad n\geq1. H t ( n ) = H t ∧ n , n ≥ 1.
Since H ( n ) H^{(n)} H ( n ) is bounded, nonnegative and optional, the preceding identity gives
E ∫ ( 0 , T ] H t ( n ) μ a ≤ ( d t ) = E ∫ ( 0 , T ] H t ( n ) ν ( d t ) . \mathbb{E}\int_{(0,T]} H_t^{(n)}\,\mu_a^{\le}(dt)=\mathbb{E}\int_{(0,T]} H_t^{(n)}\,\nu(dt). E ∫ ( 0 , T ] H t ( n ) μ a ≤ ( d t ) = E ∫ ( 0 , T ] H t ( n ) ν ( d t ) .
As n → ∞ n\to\infty n → ∞ , H t ( n ) ↑ H t H_t^{(n)}\uparrow H_t H t ( n ) ↑ H t . The monotone convergence theorem therefore yields
E ∫ ( 0 , T ] H t μ a ≤ ( d t ) = E ∫ ( 0 , T ] H t ν ( d t ) . \mathbb{E}\int_{(0,T]} H_t\,\mu_a^{\le}(dt)=\mathbb{E}\int_{(0,T]} H_t\,\nu(dt). E ∫ ( 0 , T ] H t μ a ≤ ( d t ) = E ∫ ( 0 , T ] H t ν ( d t ) .
Both integrals are nonnegative random variables. Consequently, Lemma 2.3 implies
∫ ( 0 , T ] H t μ a ≤ ( d t ) = 0 a.s. \int_{(0,T]} H_t\,\mu_a^{\le}(dt)=0\quad\text{a.s.} ∫ ( 0 , T ] H t μ a ≤ ( d t ) = 0 a.s.
if and only if
∫ ( 0 , T ] H t ν ( d t ) = 0 a.s. \int_{(0,T]} H_t\,\nu(dt)=0\quad\text{a.s.} ∫ ( 0 , T ] H t ν ( d t ) = 0 a.s.
This establishes (5).
Step 2: Identification of the projected measure.
Since a a a is an F \mathbb{F} F -optional increasing process, the optional-projection integration theorem for integration against an adapted optional increasing process applies to its associated Stieltjes measure (with the optional Doléans measure as reference).
For every bounded nonnegative F \mathbb{F} F -optional process H H H , it gives
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t o , F ( 1 { t < τ } ) d a t . \mathbb{E}\int_{(0,T]} H_t1_{\{t<\tau\}}\,da_t=\mathbb{E}\int_{(0,T]} H_t^{\mathrm{o},\mathbb{F}}\bigl(1_{\{t<\tau\}}\bigr)\,da_t. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t o , F ( 1 { t < τ } ) d a t .
By the definition of the Azéma supermartingale, o , F ( 1 { t < τ } ) = G t {}^{\mathrm{o},\mathbb{F}}\bigl(1_{\{t<\tau\}}\bigr)=G_t o , F ( 1 { t < τ } ) = G t . Hence
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t G t d a t . (7) \mathbb{E}\int_{(0,T]} H_t1_{\{t<\tau\}}\,da_t=\mathbb{E}\int_{(0,T]} H_tG_t\,da_t. \tag*{(7)} E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t G t d a t . (7)
On the other hand, the definition of ν \nu ν gives
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t ν ( d t ) . \mathbb{E}\int_{(0,T]} H_t1_{\{t<\tau\}}\,da_t=\mathbb{E}\int_{(0,T]} H_t\,\nu(dt). E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t ν ( d t ) .
Combining these identities, we obtain
E ∫ ( 0 , T ] H t ν ( d t ) = E ∫ ( 0 , T ] H t G t d a t . \mathbb{E}\int_{(0,T]} H_t\,\nu(dt)=\mathbb{E}\int_{(0,T]} H_tG_t\,da_t. E ∫ ( 0 , T ] H t ν ( d t ) = E ∫ ( 0 , T ] H t G t d a t .
Since this holds for every bounded nonnegative optional test process, uniqueness of the dual optional projection yields ν ( d t ) = G t d a t \nu(dt)=G_t\,da_t ν ( d t ) = G t d a t in the sense of optional random measures. Monotone approximation extends (7) to arbitrary nonnegative optional H H H . Applying Lemma 2.3 once again proves (6).
Step 3: Necessary and sufficient positivity condition.
Suppose first that
G > 0 m a o -a.e. G>0\quad m_a^{\mathrm{o}}\text{-a.e.} G > 0 m a o -a.e.
Assume
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. \int_{(0,T]} H_t1_{\{t<\tau\}}\,da_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s.
By (7) ,
E ∫ ( 0 , T ] H t G t d a t = 0. \mathbb{E}\int_{(0,T]} H_tG_t\,\mathrm{d}a_t=0. E ∫ ( 0 , T ] H t G t d a t = 0.
Equivalently,
∫ Ω × ( 0 , T ] H G d m a o = 0. \int_{\Omega\times(0,T]} HG\,\mathrm{d}m_a^o=0. ∫ Ω × ( 0 , T ] H G d m a o = 0.
Since H G ≥ 0 HG\ge0 H G ≥ 0 , it follows that
H G = 0 m a o -a.e. HG=0 \qquad m_a^o\text{-a.e.} H G = 0 m a o -a.e.
The strict positivity of G G G therefore implies
H = 0 m a o -a.e. H=0 \qquad m_a^o\text{-a.e.} H = 0 m a o -a.e.
Consequently,
E ∫ ( 0 , T ] H t d a t = 0 , \mathbb{E}\int_{(0,T]} H_t\,\mathrm{d}a_t=0, E ∫ ( 0 , T ] H t d a t = 0 ,
and hence
∫ ( 0 , T ] H t d a t = 0 a.s. \int_{(0,T]} H_t\,\mathrm{d}a_t=0 \qquad\text{a.s.} ∫ ( 0 , T ] H t d a t = 0 a.s.
The reverse implication follows immediately from
0 ≤ ∫ ( 0 , T ] H t 1 { t < τ } d a t ≤ ∫ ( 0 , T ] H t d a t . 0\le\int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t\le\int_{(0,T]} H_t\,\mathrm{d}a_t. 0 ≤ ∫ ( 0 , T ] H t 1 { t < τ } d a t ≤ ∫ ( 0 , T ] H t d a t .
To prove necessity, suppose that m a o ( { G = 0 } ) > 0 m_a^o(\{G=0\})>0 m a o ({ G = 0 }) > 0 . Consider the nonnegative optional process H t = 1 { G t = 0 } H_t=1_{\{G_t=0\}} H t = 1 { G t = 0 } . Then H t G t = 0 H_tG_t=0 H t G t = 0 everywhere, so
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. \int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t=0 \qquad\text{a.s.} ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s.
However,
E ∫ ( 0 , T ] H t d a t = m a o ( { G = 0 } ) > 0. \mathbb{E}\int_{(0,T]} H_t\,\mathrm{d}a_t=m_a^o(\{G=0\})>0. E ∫ ( 0 , T ] H t d a t = m a o ({ G = 0 }) > 0.
Thus the unstopped integral does not vanish almost surely. Therefore, universal equivalence of the stopped and unstopped contact conditions requires
G > 0 m a o -a.e. G>0 \qquad m_a^o\text{-a.e.} G > 0 m a o -a.e.
The proof is complete.
Remark 3.2 (Support of the reflection measure). The positivity condition
G > 0 m a o -a.e. G>0 \qquad m_a^o\text{-a.e.} G > 0 m a o -a.e.
is a condition relative to the mean measure induced by d a \mathrm{d}a d a , not to Lebesgue measure. In particular, positivity of G G G for Lebesgue-almost every time does not suffice when d a \mathrm{d}a d a has an atomic or singular component. The criterion means that the reflection measure does not charge, in the mean-measure sense, any optional region on which the conditional probability of survival is zero.
Predictable contact conditions
We now consider the case in which the increasing process a a a is F \mathbb{F} F -predictable. This distinction is particularly relevant for reflected BSDEs with irregular obstacles, where the predictable increasing component satisfies a contact condition involving predictable integrands.
Let q t = p , F ( 1 { t ≤ τ } ) q_t = {}^{p,\mathbb{F}}(\mathbf{1}_{\{t\leq\tau\}}) q t = p , F ( 1 { t ≤ τ } ) be the F \mathbb{F} F -predictable projection of the non-strict survival indicator. In terms of the Azéma supermartingale (4) , the predictable projection admits the representation
q t = G t − , t > 0 , q_t = G_{t-}, \qquad t>0, q t = G t − , t > 0 ,
with the appropriate predictable version.
We retain the notation q q q to emphasize the predictable nature of the projection and the endpoint convention.
Theorem 3.3 (Predictable contact: sharp survival criterion). Suppose, in addition, that a a a is F \mathbb{F} F -predictable. Let H ≥ 0 H\geq0 H ≥ 0 be an F \mathbb{F} F -predictable process.
Then
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t q t d a t . (8) \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t
=
\mathbb{E}\int_{(0,T]} H_tq_t\,da_t. \tag*{(8)} E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t q t d a t . (8) Consequently,
∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t q t d a t = 0 a.s. (9) \int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t=0\quad\text{a.s.}
\quad\Longleftrightarrow\quad
\int_{(0,T]} H_tq_t\,da_t=0\quad\text{a.s.} \tag*{(9)} ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t q t d a t = 0 a.s. (9) Moreover, the stopped predictable contact condition is equivalent to
∫ ( 0 , T ] H t d a t = 0 a.s. \int_{(0,T]} H_t\,da_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t d a t = 0 a.s. for every nonnegative F \mathbb{F} F -predictable process H H H if and only if
q > 0 m a p -almost everywhere , q>0\qquad m_a^p\text{-almost everywhere}, q > 0 m a p -almost everywhere , where m a p m_a^p m a p is the mean measure associated with d a da d a on P ( F ) \mathcal{P}(\mathbb{F}) P ( F ) , as introduced in Section 2.
Proof. We first establish the expectation identity (8) . Since a a a is an F \mathbb{F} F -predictable increasing process, the predictable-projection integration theorem gives, for every bounded nonnegative predictable process H H H ,
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] p , F ( H t 1 { t ≤ τ } ) d a t . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t
=
\mathbb{E}\int_{(0,T]} {}^{p,\mathbb{F}}\left(H_t\mathbf{1}_{\{t\leq\tau\}}\right)\,da_t. E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] p , F ( H t 1 { t ≤ τ } ) d a t .
Because H H H is predictable, the multiplication property of predictable projection yields p , F ( H t 1 { t ≤ τ } ) = H t q t {}^{p,\mathbb{F}}\left(H_t\mathbf{1}_{\{t\leq\tau\}}\right)=H_tq_t p , F ( H t 1 { t ≤ τ } ) = H t q t . Therefore,
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t q t d a t . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t
=
\mathbb{E}\int_{(0,T]} H_tq_t\,da_t. E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t q t d a t .
For an arbitrary nonnegative predictable H H H , apply this identity to H ( n ) = H ∧ n H^{(n)}=H\wedge n H ( n ) = H ∧ n and let n → ∞ n\to\infty n → ∞ . The monotone convergence theorem proves (8) . Since both integrals in (8) are nonnegative random variables, Lemma 2.3 gives (9) .
We now prove the positivity criterion.
Suppose first that
q > 0 m a p -a.e. q>0\qquad m_a^p\text{-a.e.} q > 0 m a p -a.e.
If the stopped contact condition holds, then
E ∫ ( 0 , T ] H t q t d a t = 0. \mathbb{E}\int_{(0,T]} H_tq_t\,da_t = 0. E ∫ ( 0 , T ] H t q t d a t = 0.
By the definition of the mean measure,
∫ Ω × ( 0 , T ] H q d m a p = 0. \int_{\Omega\times(0,T]} Hq\,dm_a^p = 0. ∫ Ω × ( 0 , T ] H q d m a p = 0.
Since H q ≥ 0 Hq \ge0 H q ≥ 0 , this implies
H q = 0 m a p -a.e. Hq = 0 \qquad m_a^p\text{-a.e.} H q = 0 m a p -a.e.
Using the strict positivity of q q q , we conclude that
H = 0 m a p -a.e. H = 0 \qquad m_a^p\text{-a.e.} H = 0 m a p -a.e.
Thus
E ∫ ( 0 , T ] H t d a t = 0 , \mathbb{E}\int_{(0,T]} H_t\,da_t = 0, E ∫ ( 0 , T ] H t d a t = 0 ,
and consequently
∫ ( 0 , T ] H t d a t = 0 a.s. \int_{(0,T]} H_t\,da_t = 0 \qquad\text{a.s.} ∫ ( 0 , T ] H t d a t = 0 a.s.
Conversely, if
∫ ( 0 , T ] H t d a t = 0 a.s., \int_{(0,T]} H_t\,da_t = 0 \qquad\text{a.s.,} ∫ ( 0 , T ] H t d a t = 0 a.s.,
then, by nonnegativity,
∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s. \int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t = 0 \qquad\text{a.s.} ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s.
This proves sufficiency.
For necessity, suppose that
m a p ( { q = 0 } ) > 0. m_a^p(\{q=0\}) > 0. m a p ({ q = 0 }) > 0.
Choose
H t = 1 { q t = 0 } . H_t = \mathbf{1}_{\{q_t=0\}}. H t = 1 { q t = 0 } .
This process is nonnegative and predictable. Moreover, H t q t = 0 H_tq_t=0 H t q t = 0 . By (8),
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 , \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t = 0, E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 ,
and hence
∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s. \int_{(0,T]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t = 0 \qquad\text{a.s.} ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = 0 a.s.
On the other hand,
E ∫ ( 0 , T ] H t d a t = m a p ( { q = 0 } ) > 0. \mathbb{E}\int_{(0,T]} H_t\,da_t = m_a^p(\{q=0\}) > 0. E ∫ ( 0 , T ] H t d a t = m a p ({ q = 0 }) > 0.
Therefore, the unstopped contact condition does not hold almost surely. The equivalence for every nonnegative predictable H H H is thus possible only if
q > 0 m a p -a.e. q > 0 \qquad m_a^p\text{-a.e.} q > 0 m a p -a.e.
This completes the proof.
Examples and interpretation
The following examples illustrate why survival positivity and endpoint conventions are essential.
Example 3.5 (An atom at the random horizon). Let T = 2 T=2 T = 2 , let τ = 1 \tau=1 τ = 1 deterministically, and consider a t = 1 { t ≥ 1 } a_t=\mathbf{1}_{\{t\geq1\}} a t = 1 { t ≥ 1 } .
The associated Stieltjes measure is d a t = δ 1 ( d t ) da_t=\delta_1(dt) d a t = δ 1 ( d t ) (see Table 1 ).
For H t ≡ 1 H_t\equiv1 H t ≡ 1 , we have
∫ ( 0 , 2 ] H t 1 { t < τ } d a t = 0. \int_{(0,2]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t=0. ∫ ( 0 , 2 ] H t 1 { t < τ } d a t = 0. whereas
∫ ( 0 , 2 ] H t 1 { t ≤ τ } d a t = 1. \int_{(0,2]} H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t=1. ∫ ( 0 , 2 ] H t 1 { t ≤ τ } d a t = 1. Thus, the two stopped random measures assign different masses to the graph of the random horizon. This example shows that the distinction between strict and non-strict stopping cannot be ignored for atomic reflection measures.
Table 1. Examples of increasing processes and their associated Stieltjes measures.
Example 3.6 (Failure of contact transfer without survival positivity). Let T = 2 T=2 T = 2 , let τ = 1 \tau=1 τ = 1 deterministically, and define
a t = ( t − 1 ) + , H t ≡ 1. a_t=(t-1)^{+}, \qquad H_t\equiv1. a t = ( t − 1 ) + , H t ≡ 1. Then
d a t = 1 { t > 1 } d t . da_t=\mathbf{1}_{\{t>1\}}\,dt. d a t = 1 { t > 1 } d t . Since d a da d a is supported on ( 1 , 2 ] (1,2] ( 1 , 2 ] , we obtain
∫ ( 0 , 2 ] H t 1 { t < τ } d a t = 0. \int_{(0,2]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t=0. ∫ ( 0 , 2 ] H t 1 { t < τ } d a t = 0. However,
∫ ( 0 , 2 ] H t d a t = 1. \int_{(0,2]} H_t\,da_t=1. ∫ ( 0 , 2 ] H t d a t = 1. The reference filtration may be taken to be trivial. In that case,
G t = 1 { t < 1 } , G_t=\mathbf{1}_{\{t<1\}}, G t = 1 { t < 1 } , which vanishes on the support of d a da d a .
Thus the stopped contact condition holds, while the unstopped contact condition fails. The example demonstrates that the positivity condition must be imposed relative to the reflection measure.
The Example 3.6 is intended to illustrate the measure-theoretic obstruction only; it is not a counterexample to any specific invariance-time hypothesis.
Cox construction: strong survival equivalence
We now specialize the preceding contact-transfer criteria to a Cox construction. In this setting, the conditional survival process is explicit and strictly positive on the finite horizon. This allows us to establish a strong equivalence between stopped and unstopped contact conditions, without imposing absolute continuity of the underlying reflection measure.
Cox random time and conditional survival
Assume that the probability space carries a nonnegative, continuous, nondecreasing, F \mathbb{F} F -adapted process Λ = ( Λ t ) t ≥ 0 \Lambda=(\Lambda_t)_{t\geq0} Λ = ( Λ t ) t ≥ 0 satisfying
Λ 0 = 0 , Λ T < ∞ a.s. \Lambda_0=0,\qquad\Lambda_T<\infty\quad\text{a.s.} Λ 0 = 0 , Λ T < ∞ a.s.
Let E E E be a unit exponential random variable ( E ∼ Exp ( 1 ) ) (E\sim\operatorname{Exp}(1)) ( E ∼ Exp ( 1 )) , independent of F ∞ \mathcal{F}_{\infty} F ∞ , and define the Cox random time
τ = inf { t ≥ 0 : Λ t ≥ E } , (10) \tau=\inf\{t\geq0:\Lambda_t\geq E\}, \tag*{(10)} τ = inf { t ≥ 0 : Λ t ≥ E } , (10)
with the convention inf ∅ = ∞ \inf\varnothing=\infty inf ∅ = ∞ .
The process Λ \Lambda Λ is called the cumulative hazard process . When it is absolutely continuous, it admits a representation
Λ t = ∫ 0 t λ s d s , \Lambda_t=\int_0^t\lambda_s\,\mathrm{d}s, Λ t = ∫ 0 t λ s d s ,
where λ \lambda λ is a nonnegative F \mathbb{F} F -progressively measurable process, called the hazard-rate or intensity process .
Define the survival process
S t = e − Λ t , 0 ≤ t ≤ T . S_t=e^{-\Lambda_t},\qquad0\leq t\leq T. S t = e − Λ t , 0 ≤ t ≤ T .
By continuity and monotonicity of Λ \Lambda Λ ,
{ τ > t } = { E > Λ t } . \{\tau> t\} = \{E > \Lambda_t\}. { τ > t } = { E > Λ t } .
Since E E E is independent of F ∞ \mathcal{F}_\infty F ∞ and has the unit exponential distribution,
P ( τ > t ∣ F ∞ ) = P ( E > Λ t ∣ F ∞ ) = e − Λ t = S t . \mathbb{P}(\tau> t \mid\mathcal{F}_\infty) = \mathbb{P}(E > \Lambda_t \mid\mathcal{F}_\infty) = e^{-\Lambda_t} = S_t. P ( τ > t ∣ F ∞ ) = P ( E > Λ t ∣ F ∞ ) = e − Λ t = S t .
Furthermore, since Λ t \Lambda_t Λ t is F t \mathcal{F}_t F t -measurable, the tower property gives
G t = P ( τ > t ∣ F t ) = S t = e − Λ t . G_t = \mathbb{P}(\tau> t \mid\mathcal{F}_t) = S_t = e^{-\Lambda_t}. G t = P ( τ > t ∣ F t ) = S t = e − Λ t .
Thus, the Azéma supermartingale associated with τ \tau τ is explicitly identified. In particular, the finiteness of Λ T \Lambda_T Λ T implies
S t > 0 , 0 ≤ t ≤ T , a.s. (11) S_t > 0,\qquad0 \le t \le T,\qquad\text{a.s.} \tag*{(11)} S t > 0 , 0 ≤ t ≤ T , a.s. (11)
This strict positivity (11) is the key property underlying the equivalence of contact conditions.
Remark 4.1 (Financial interpretation). In reduced-form credit-risk modeling, τ \tau τ may represent the default time of a counterparty, while Λ t \Lambda_t Λ t describes the cumulative default hazard.
The process S t = e − Λ t = G t S_t = e^{-\Lambda_t} = G_t S t = e − Λ t = G t represents the probability of survival beyond time t t t , conditional on the reference information.
In the particular case of a constant intensity λ > 0 \lambda> 0 λ > 0 , using (10), we get
Λ t = λ t , τ = E λ , \Lambda_t = \lambda t,\qquad\tau= \frac{E}{\lambda}, Λ t = λ t , τ = λ E ,
and hence
P ( τ > t ) = e − λ t . \mathbb{P}(\tau> t) = e^{-\lambda t}. P ( τ > t ) = e − λ t .
Thus, τ \tau τ follows an exponential distribution with parameter λ \lambda λ .
The Cox construction provides a tractable model in which default information can be introduced through progressive enlargement, while the reference filtration retains the dynamics of the underlying market factors.
Conditional survival identity for random measures
The following result is the central measure-theoretic identity of this section. Its significance is that it applies to arbitrary integrable Stieltjes measures associated with the reference filtration, including measures with atoms or singular continuous components.
Lemma 4.2 (Conditional Fubini for Cox survival). Let H ≥ 0 H \ge0 H ≥ 0 be an F \mathbb{F} F -optional process, and let a a a satisfy the standing assumptions of Section 2.
Then
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . (12) \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t
=
\mathbb{E}\int_{(0,T]} H_tS_t\,da_t. \tag*{(12)} E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . (12) Moreover, the same identity holds for the non-strict survival indicator:
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t S t d a t . (13) \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\le\tau\}}\,da_t
=
\mathbb{E}\int_{(0,T]} H_tS_t\,da_t. \tag*{(13)} E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t S t d a t . (13) Consequently,
∫ ( 0 , T ] H t 1 { t < τ } d a t = ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t a.s. (14) \int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t
=
\int_{(0,T]} H_t\mathbf{1}_{\{t\le\tau\}}\,da_t
\quad\text{a.s.} \tag*{(14)} ∫ ( 0 , T ] H t 1 { t < τ } d a t = ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t a.s. (14)
Proof. We first establish the strict survival identity (12).
Since H H H , a a a , and Λ \Lambda Λ are F \mathbb{F} F -optional or adapted, their restrictions to [ 0 , T ] [0,T] [ 0 , T ] are measurable with respect to the reference information F ∞ \mathcal{F}_{\infty} F ∞ .
Conditionally on F ∞ \mathcal{F}_{\infty} F ∞ , the paths of H H H , a a a , and Λ \Lambda Λ are fixed, whereas the exponential random variable E E E remains independent.
By the definition of the Cox time, we have 1 { t < τ } = 1 { E > Λ t } \mathbf{1}_{\{t<\tau\}}=\mathbf{1}_{\{E>\Lambda_t\}} 1 { t < τ } = 1 { E > Λ t } .
For every fixed t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , independence yields
E [ 1 { E > Λ t } ∣ F ∞ ] = P ( E > Λ t ∣ F ∞ ) = e − Λ t = S t . \mathbb{E}\left[\mathbf{1}_{\{E>\Lambda_t\}}\mid\mathcal{F}_{\infty}\right]
=\mathbb{P}(E>\Lambda_t\mid\mathcal{F}_{\infty})
=e^{-\Lambda_t}=S_t. E [ 1 { E > Λ t } ∣ F ∞ ] = P ( E > Λ t ∣ F ∞ ) = e − Λ t = S t .
Because H ≥ 0 H\geq0 H ≥ 0 and d a da d a is a positive random measure, conditional Tonelli’s theorem allows us to interchange conditional expectation and integration with respect to the fixed Stieltjes measure. Consequently,
E [ ∫ ( 0 , T ] H t 1 { t < τ } d a t ∣ F ∞ ] = ∫ ( 0 , T ] H t E [ 1 { E > Λ t } ∣ F ∞ ] d a t = ∫ ( 0 , T ] H t e − Λ t d a t = ∫ ( 0 , T ] H t S t d a t . \mathbb{E}\left[\left.\int_{(0,T]}H_t\mathbf{1}_{\{t<\tau\}}\,da_t\right|\mathcal{F}_{\infty}\right]
=\int_{(0,T]}H_t\mathbb{E}\left[\mathbf{1}_{\{E>\Lambda_t\}}\mid\mathcal{F}_{\infty}\right]\,da_t
=\int_{(0,T]}H_te^{-\Lambda_t}\,da_t
=\int_{(0,T]}H_tS_t\,da_t. E [ ∫ ( 0 , T ] H t 1 { t < τ } d a t F ∞ ] = ∫ ( 0 , T ] H t E [ 1 { E > Λ t } ∣ F ∞ ] d a t = ∫ ( 0 , T ] H t e − Λ t d a t = ∫ ( 0 , T ] H t S t d a t .
Taking expectations gives
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . \mathbb{E}\int_{(0,T]}H_t\mathbf{1}_{\{t<\tau\}}\,da_t
=\mathbb{E}\int_{(0,T]}H_tS_t\,da_t. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t .
which proves (12).
We next consider the non-strict survival indicator.
By continuity of Λ \Lambda Λ , the events { t ≤ τ } \{t\leq\tau\} { t ≤ τ } and { E ≥ Λ t } \{E\geq\Lambda_t\} { E ≥ Λ t } coincide up to the possible equality event { E = Λ t } \{E=\Lambda_t\} { E = Λ t } .
Since the exponential distribution has no atoms, P ( E = Λ t ∣ F ∞ ) = 0 \mathbb{P}(E=\Lambda_t\mid\mathcal{F}_{\infty})=0 P ( E = Λ t ∣ F ∞ ) = 0 for every fixed t t t . The same conditional Tonelli argument therefore gives
E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t S t d a t . \mathbb{E}\int_{(0,T]}H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t
=\mathbb{E}\int_{(0,T]}H_tS_t\,da_t. E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = E ∫ ( 0 , T ] H t S t d a t .
proving (13).
Finally, the difference between the two stopped integrals is, by the boundary identity established in (3) ,
∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = ∫ ( 0 , T ] H t 1 { t < τ } d a t + H τ Δ a τ 1 { 0 < τ ≤ T } . \int_{(0,T]}H_t\mathbf{1}_{\{t\leq\tau\}}\,da_t
=\int_{(0,T]}H_t\mathbf{1}_{\{t<\tau\}}\,da_t
+H_{\tau}\Delta a_{\tau}\mathbf{1}_{\{0<\tau\leq T\}}. ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t = ∫ ( 0 , T ] H t 1 { t < τ } d a t + H τ Δ a τ 1 { 0 < τ ≤ T } .
The left-hand side is nonnegative, and the preceding identities show that its expectation is zero.
It follows that
H τ Δ a τ 1 { 0 < τ ≤ T } = 0 a.s. H_{\tau}\Delta a_{\tau}\mathbf{1}_{\{0<\tau\leq T\}}=0\quad\text{a.s.} H τ Δ a τ 1 { 0 < τ ≤ T } = 0 a.s.
Hence the strict and non-strict stopped integrals coincide almost surely, which proves (14). □ \square □
Remark 4.3 (Why atoms do not cause a boundary problem). The preceding result does not require continuity of the increasing process ( a t ) (a_t) ( a t ) . Indeed, ( d a t ) (da_t) ( d a t ) may have atoms at F \mathbb{F} F -stopping times or may possess a singular continuous component. The essential feature is that the Cox threshold E E E is independent of the reference information and has a continuous distribution.
Conditionally on F ∞ \mathcal{F}_{\infty} F ∞ , the Stieltjes measure d a \mathrm{d}a d a and the cumulative hazard Λ \Lambda Λ are fixed. Thus, even when ( d a t ) (\mathrm{d}a_t) ( d a t ) has atoms, the probability that the Cox random time coincides with an atom of ( d a t ) (\mathrm{d}a_t) ( d a t ) is zero. More precisely,
Δ a τ 1 { 0 < τ ≤ T } = 0 a.s. \Delta a_{\tau}1_{\{0<\tau\le T\}}=0 \quad\text{a.s.} Δ a τ 1 { 0 < τ ≤ T } = 0 a.s.
This property is specific to the continuous-threshold Cox construction and need not hold for a general random time.
Strong equivalence of contact conditions
We can now establish the transfer of contact conditions without imposing absolute continuity of the reflection measure.
Corollary 4.4 (Cox contact transfer for arbitrary Stieltjes reflection measures). Under the assumptions of Lemma 4.2, for every nonnegative F \mathbb{F} F -optional process H H H ,
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t d a t = 0 a.s. (15) \int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t=0 \quad\text{a.s.}
\quad\Longleftrightarrow\quad
\int_{(0,T]} H_t\,\mathrm{d}a_t=0 \quad\text{a.s.} \tag*{(15)} ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t d a t = 0 a.s. (15) The same equivalence holds with 1 { t ≤ τ } 1_{\{t\le\tau\}} 1 { t ≤ τ } in place of 1 { t < τ } 1_{\{t<\tau\}} 1 { t < τ } . In particular, the result applies to F \mathbb{F} F -predictable increasing processes and predictable contact integrands, as well as to random Stieltjes measures with atomic or singular continuous components.
Proof. By Lemma 4.2,
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . \mathbb{E}\int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t
=
\mathbb{E}\int_{(0,T]} H_tS_t\,\mathrm{d}a_t. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t .
Since both integrals are nonnegative, Lemma 2.3 implies
∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t S t d a t = 0 a.s. \int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t=0 \quad\text{a.s.}
\quad\Longleftrightarrow\quad
\int_{(0,T]} H_tS_t\,\mathrm{d}a_t=0 \quad\text{a.s.} ∫ ( 0 , T ] H t 1 { t < τ } d a t = 0 a.s. ⟺ ∫ ( 0 , T ] H t S t d a t = 0 a.s.
Now, by (11), S t = e − Λ t > 0 S_t=e^{-\Lambda_t}>0 S t = e − Λ t > 0 for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , almost surely. Since Λ \Lambda Λ is nonnegative and nondecreasing, we have
0 ≤ Λ t ≤ Λ T < ∞ , 0 ≤ t ≤ T , 0\le\Lambda_t\le\Lambda_T<\infty,\qquad0\le t\le T, 0 ≤ Λ t ≤ Λ T < ∞ , 0 ≤ t ≤ T ,
almost surely. Therefore, 0 < e − Λ T ≤ S t = e − Λ t ≤ 1 0<e^{-\Lambda_T}\le S_t=e^{-\Lambda_t}\le1 0 < e − Λ T ≤ S t = e − Λ t ≤ 1 . Since H t ≥ 0 H_t\ge0 H t ≥ 0 and d a t \mathrm{d}a_t d a t is a positive measure, integration yields
e − Λ T ∫ ( 0 , T ] H t d a t ≤ ∫ ( 0 , T ] H t S t d a t ≤ ∫ ( 0 , T ] H t d a t . e^{-\Lambda_T}\int_{(0,T]} H_t\,\mathrm{d}a_t
\le
\int_{(0,T]} H_tS_t\,\mathrm{d}a_t
\le
\int_{(0,T]} H_t\,\mathrm{d}a_t. e − Λ T ∫ ( 0 , T ] H t d a t ≤ ∫ ( 0 , T ] H t S t d a t ≤ ∫ ( 0 , T ] H t d a t .
Since e − Λ T > 0 e^{-\Lambda_T}>0 e − Λ T > 0 almost surely, the middle integral vanishes if and only if the right-hand integral vanishes. Hence,
∫ ( 0 , T ] H t S t d a t = 0 ⟺ ∫ ( 0 , T ] H t d a t = 0 , \int_{(0,T]} H_tS_t\,\mathrm{d}a_t=0
\quad\Longleftrightarrow\quad
\int_{(0,T]} H_t\,\mathrm{d}a_t=0, ∫ ( 0 , T ] H t S t d a t = 0 ⟺ ∫ ( 0 , T ] H t d a t = 0 ,
almost surely. Combining the two equivalences proves (15). The non-strict case follows from (13) by the same argument. □ \square □
Consequently, the strict positivity condition
G > 0 m a o -a.e. G > 0 \quad m_a^{\mathrm{o}}\text{-a.e.} G > 0 m a o -a.e.
is automatically satisfied for every integrable F \mathbb{F} F -adapted increasing process a a a . Thus, the Cox construction ensures that stopping at τ \tau τ does not destroy contact information carried by any integrable reference Stieltjes measure.
Example 4.6. (Contact transfer for an atomic measure). Let T > t 0 > 0 T > t_0 > 0 T > t 0 > 0 , and suppose that
Λ t = λ t , λ > 0. \Lambda_t = \lambda t, \qquad\lambda> 0. Λ t = λ t , λ > 0. Then
τ = E λ , S t = e − λ t . \tau= \frac{E}{\lambda}, \qquad S_t = e^{-\lambda t}. τ = λ E , S t = e − λ t . Consider the increasing process a t = 1 { t ≥ t 0 } a_t = \mathbf{1}_{\{t \ge t_0\}} a t = 1 { t ≥ t 0 } , whose Stieltjes measure is d a t = δ t 0 ( d t ) da_t = \delta_{t_0}(dt) d a t = δ t 0 ( d t ) . For any nonnegative deterministic function H H H , we obtain
∫ ( 0 , T ] H t 1 { t < τ } d a t = H t 0 1 { t 0 < τ } . \int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t = H_{t_0}\mathbf{1}_{\{t_0<\tau\}}. ∫ ( 0 , T ] H t 1 { t < τ } d a t = H t 0 1 { t 0 < τ } . Taking expectations gives
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = H t 0 P ( τ > t 0 ) = H t 0 e − λ t 0 . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t
= H_{t_0}\mathbb{P}(\tau>t_0)
= H_{t_0}e^{-\lambda t_0}. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = H t 0 P ( τ > t 0 ) = H t 0 e − λ t 0 . On the other hand,
∫ ( 0 , T ] H t S t d a t = H t 0 e − λ t 0 . \int_{(0,T]} H_tS_t\,da_t = H_{t_0}e^{-\lambda t_0}. ∫ ( 0 , T ] H t S t d a t = H t 0 e − λ t 0 . Therefore,
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t
= \mathbb{E}\int_{(0,T]} H_tS_t\,da_t. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t S t d a t . Since e − λ t 0 > 0 e^{-\lambda t_0}>0 e − λ t 0 > 0 , the stopped contact condition is equivalent to H t 0 = 0 H_{t_0}=0 H t 0 = 0 , which is precisely the unstopped contact condition
∫ ( 0 , T ] H t d a t = 0. \int_{(0,T]} H_t\,da_t = 0. ∫ ( 0 , T ] H t d a t = 0. This example shows explicitly that the transfer result remains valid when the entire reflection measure is concentrated at a single time.
The results of this section concern the preservation of zero-contact conditions under Cox stopping. They do not, by themselves, establish the reduction or reconstruction of a reflected backward stochastic differential equation. A complete reconstruction theorem additionally requires appropriate martingale representations, a compatible default compensator, a transformed generator, terminal conditions, and obstacle regularity. These additional issues will be addressed in the next section for a Brownian Cox model with continuous reflection.
Extension to reference-measurable random measures
The conditional Tonelli argument of Lemma 4.2 continues to hold when the Stieltjes measure d a da d a is replaced by an arbitrary finite random measure determined by the reference information. We record this extension without repeating the preceding proof.
Lemma 4.7 (Cox weighting beyond adapted Stieltjes measures). Let μ \mu μ be a finite random measure on ( 0 , T ] (0,T] ( 0 , T ] whose values on Borel sets are F ∞ \mathcal{F}_{\infty} F ∞ -measurable, and let H ≥ 0 H \ge0 H ≥ 0 be jointly F ∞ ⊗ B ( ( 0 , T ] ) \mathcal{F}_{\infty} \otimes\mathcal{B}((0,T]) F ∞ ⊗ B (( 0 , T ]) -measurable. Under the Cox assumptions of § 4 ,
E ∫ ( 0 , T ] H t 1 { t < τ } μ ( d t ) = E ∫ ( 0 , T ] H t e − Λ t μ ( d t ) = E ∫ ( 0 , T ] H t 1 { t ≤ τ } μ ( d t ) . \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,\mu(\mathrm{d}t) = \mathbb{E}\int_{(0,T]} H_t e^{-\Lambda_t}\,\mu(\mathrm{d}t) = \mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t\le\tau\}}\,\mu(\mathrm{d}t). E ∫ ( 0 , T ] H t 1 { t < τ } μ ( d t ) = E ∫ ( 0 , T ] H t e − Λ t μ ( d t ) = E ∫ ( 0 , T ] H t 1 { t ≤ τ } μ ( d t ) . The identities are valid for atomic, singular continuous, and absolutely continuous random measures alike.
Proof. Condition on F ∞ \mathcal{F}_{\infty} F ∞ . The measure μ \mu μ and processes H , Λ H,\Lambda H , Λ are then fixed, while E E E retains its unit exponential law. The conditional probabilities of { E > Λ t } \{E>\Lambda_t\} { E > Λ t } and { E ≥ Λ t } \{E\ge\Lambda_t\} { E ≥ Λ t } are both e − Λ t e^{-\Lambda_t} e − Λ t , because the exponential distribution is atomless. Conditional Tonelli’s theorem and then the tower property give the stated identities.
Corollary 4.8 (Zero-contact transfer for reference measures). Under the hypotheses of Lemma 4.7,
∫ ( 0 , T ] H t μ ( d t ) = 0 a.s. \int_{(0,T]} H_t\,\mu(\mathrm{d}t)=0 \quad\text{a.s.} ∫ ( 0 , T ] H t μ ( d t ) = 0 a.s. if and only if either stopped contact integral, with 1 { t < τ } \mathbf{1}_{\{t<\tau\}} 1 { t < τ } or 1 { t ≤ τ } \mathbf{1}_{\{t\le\tau\}} 1 { t ≤ τ } , vanishes almost surely.
Proof. By Lemma 4.7, the stopped integrals have the same expectation as the nonnegative integral weighted by e − Λ t e^{-\Lambda_t} e − Λ t . Since e − Λ T ≤ e − Λ t ≤ 1 e^{-\Lambda_T}\le e^{-\Lambda_t}\le1 e − Λ T ≤ e − Λ t ≤ 1 and e − Λ T > 0 e^{-\Lambda_T}>0 e − Λ T > 0 almost surely, that weighted integral vanishes precisely when the unweighted one does. The conclusion follows from nonnegativity.
A complete Brownian Cox RBSDE theorem
We now establish a reduction–reconstruction result for reflected backward stochastic differential equations under a Cox random horizon.
In contrast to the preceding sections, which concerned the transfer of contact conditions, the present result provides an explicit reconstruction of the full enlarged-filtration equation, including its default martingale, jump coefficient, and reflection process.
Probabilistic setting and assumptions
Assume that F \mathbb{F} F is the augmented natural filtration of a d d d -dimensional Brownian motion W = ( W t ) 0 ≤ t ≤ T W=(W_t)_{0\le t\le T} W = ( W t ) 0 ≤ t ≤ T . We consider the Cox construction introduced in § 4 , with
Λ t = ∫ 0 t λ s d s . \Lambda_t=\int_0^t \lambda_s\,\mathrm{d}s. Λ t = ∫ 0 t λ s d s .
where λ \lambda λ is nonnegative, F \mathbb{F} F -progressively measurable, and satisfies
0 ≤ λ t ≤ λ ˉ (16) 0\le\lambda_t\le\bar{\lambda} \tag*{(16)} 0 ≤ λ t ≤ λ ˉ (16)
for some deterministic constant λ ˉ < ∞ \bar{\lambda}<\infty λ ˉ < ∞ , d t ⊗ d P \mathrm{d}t\otimes\mathrm{d}\mathbb{P} d t ⊗ d P -almost everywhere.
The random time τ \tau τ is defined by
τ = inf { t ≥ 0 : Λ t ≥ E } , \tau=\inf\{t\ge0:\Lambda_t\ge E\}, τ = inf { t ≥ 0 : Λ t ≥ E } ,
where E E E is a unit exponential random variable independent of F ∞ \mathcal{F}_{\infty} F ∞ .
Let G \mathbb{G} G be the usual progressive enlargement of F \mathbb{F} F by τ \tau τ .
Under the Cox construction, the immersion property, also known as hypothesis ( H ) (H) ( H ) , holds between F \mathbb{F} F and its progressive enlargement G \mathbb{G} G by τ \tau τ (see Jeanblanc and Le Cam [22, Section 5]).
Consequently, every F \mathbb{F} F -local martingale remains a G \mathbb{G} G -local martingale. In particular, since W W W is continuous and ⟨ W i , W j ⟩ t = δ i j t \langle W^{i}, W^{j}\rangle_t = \delta_{ij}t ⟨ W i , W j ⟩ t = δ ij t , Lévy’s characterization implies that W W W remains a d d d -dimensional G \mathbb{G} G -Brownian motion. Define the default indicator D t = 1 { τ ≤ t } D_t = 1_{\{\tau\leq t\}} D t = 1 { τ ≤ t } , and the compensated default martingale
M t = D t − ∫ 0 t ∧ τ λ s d s . M_t = D_t - \int_0^{t\wedge\tau} \lambda_s\,ds. M t = D t − ∫ 0 t ∧ τ λ s d s .
Then M M M is a G \mathbb{G} G -martingale, and Δ M τ = 1 \Delta M_\tau= 1 Δ M τ = 1 on { τ < ∞ } \{\tau< \infty\} { τ < ∞ } . More precisely,
d D t = d M t + 1 { t ≤ τ } λ t d t , (17) dD_t = dM_t + 1_{\{t\leq\tau\}}\lambda_t\,dt, \tag*{(17)} d D t = d M t + 1 { t ≤ τ } λ t d t , (17)
where the survival indicator is understood in its predictable version.
We use the classical spaces S F 2 S_{\mathbb{F}}^2 S F 2 , H F 2 H_{\mathbb{F}}^2 H F 2 , and A F 2 A_{\mathbb{F}}^2 A F 2 .
Here, S F 2 S_{\mathbb{F}}^2 S F 2 denotes the space of continuous adapted processes y y y such that
E [ sup 0 ≤ t ≤ T ∣ y t ∣ 2 ] < ∞ , \mathbb{E}\left[\sup_{0\leq t\leq T}|y_t|^2\right] < \infty, E [ 0 ≤ t ≤ T sup ∣ y t ∣ 2 ] < ∞ ,
and H F 2 H_{\mathbb{F}}^2 H F 2 denotes the space of predictable processes z z z such that
E ∫ 0 T ∣ z t ∣ 2 d t < ∞ . \mathbb{E}\int_0^T |z_t|^2\,dt < \infty. E ∫ 0 T ∣ z t ∣ 2 d t < ∞.
The space A F 2 A_{\mathbb{F}}^2 A F 2 consists of continuous, adapted, nondecreasing processes k k k , starting from zero, satisfying
E [ k T 2 ] < ∞ . \mathbb{E}[k_T^2] < \infty. E [ k T 2 ] < ∞.
We impose the following assumptions.
(H1) Terminal condition. The terminal random variable satisfies ξ ∈ L 2 ( F T ) \xi\in L^2(\mathcal{F}_T) ξ ∈ L 2 ( F T ) .
(H2) Obstacle and recovery process. The processes L L L and R R R are continuous, F \mathbb{F} F -adapted, and belong to S F 2 S_{\mathbb{F}}^2 S F 2 . Moreover, L T ≤ ξ L_T \leq\xi L T ≤ ξ a.s. We also assume
L t ≤ R t , 0 ≤ t ≤ T , a.s. L_t \leq R_t,\qquad0\leq t\leq T,\qquad\text{a.s.} L t ≤ R t , 0 ≤ t ≤ T , a.s.
(H3) Generator. The function f : Ω × [ 0 , T ] × R × R d × R ⟶ R f:\Omega\times[0,T]\times\mathbb{R}\times\mathbb{R}^d\times\mathbb{R}\longrightarrow\mathbb{R} f : Ω × [ 0 , T ] × R × R d × R ⟶ R is progressively measurable with respect to F \mathbb{F} F for each fixed ( y , z , u ) (y,z,u) ( y , z , u ) .
There exists a constant C f > 0 C_f>0 C f > 0 such that
∣ f ( t , y , z , u ) − f ( t , y ′ , z ′ , u ′ ) ∣ ≤ C f ( ∣ y − y ′ ∣ + ∣ z − z ′ ∣ + ∣ u − u ′ ∣ ) |f(t,y,z,u)-f(t,y',z',u')| \leq C_f\left(|y-y'|+|z-z'|+|u-u'|\right) ∣ f ( t , y , z , u ) − f ( t , y ′ , z ′ , u ′ ) ∣ ≤ C f ( ∣ y − y ′ ∣ + ∣ z − z ′ ∣ + ∣ u − u ′ ∣ )
for all admissible arguments, d t ⊗ d P dt\otimes d\mathbb{P} d t ⊗ d P -almost everywhere. Furthermore,
E ∫ 0 T ∣ f ( t , 0 , 0 , 0 ) ∣ 2 d t < ∞ . \mathbb{E}\int_0^T |f(t,0,0,0)|^2\,dt < \infty. E ∫ 0 T ∣ f ( t , 0 , 0 , 0 ) ∣ 2 d t < ∞.
Define the reduced generator by
F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) . (18) F(t,y,z) = f(t,y,z,R_t-y)+\lambda_t(R_t-y). \tag*{(18)} F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) . (18)
By the Lipschitz property of f f f , the boundedness of λ \lambda λ (16) , and the square integrability of R R R , the generator F F F is progressively measurable, uniformly Lipschitz in ( y , z ) (y,z) ( y , z ) , and satisfies
E ∫ 0 T ∣ F ( t , 0 , 0 ) ∣ 2 d t < ∞ . \mathbb{E}\int_0^T |F(t,0,0)|^2\,dt < \infty. E ∫ 0 T ∣ F ( t , 0 , 0 ) ∣ 2 d t < ∞.
Thus, the standard assumptions for the Brownian reflected BSDE with continuous obstacle L L L are satisfied (see, e.g., [13 ] ).
Reduction and reconstruction theorem
Theorem 5.1 (Reduction and reconstruction with continuous reflection). Under assumptions (H1)–(H3), the following assertions hold.
(i) Existence and uniqueness of the reduced RBSDE.
There exists a unique triple
( y , z , k ) ∈ S F 2 × H F 2 × A F 2 (y,z,k) \in\mathcal{S}^{2}_{F} \times\mathcal{H}^{2}_{F} \times\mathcal{A}^{2}_{F} ( y , z , k ) ∈ S F 2 × H F 2 × A F 2 satisfying, for every t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] ,
y t = ξ + ∫ t T F ( s , y s , z s ) d s + k T − k t − ∫ t T z s d W s , (19) y_t = \xi+ \int_t^T F(s,y_s,z_s)\,\mathrm{d}s + k_T-k_t-\int_t^T z_s\,\mathrm{d}W_s, \tag*{(19)} y t = ξ + ∫ t T F ( s , y s , z s ) d s + k T − k t − ∫ t T z s d W s , (19) together with
y t ≥ L t , 0 ≤ t ≤ T , (20) y_t \ge L_t,\qquad0 \le t \le T, \tag*{(20)} y t ≥ L t , 0 ≤ t ≤ T , (20) and
∫ ( 0 , T ] ( y t − L t ) d k t = 0. (21) \int_{(0,T]} (y_t-L_t)\,\mathrm{d}k_t = 0. \tag*{(21)} ∫ ( 0 , T ] ( y t − L t ) d k t = 0. (21) (ii) Reconstruction in the enlarged filtration.
Define, for 0 ≤ t ≤ T 0 \le t \le T 0 ≤ t ≤ T ,
Y t = y t 1 { t < τ } + R τ 1 { τ ≤ t } , Z t = z t 1 { t ≤ τ } , U t = ( R t − y t ) 1 { t ≤ τ } , K t = k t ∧ τ . (22) \begin{aligned}
Y_t &= y_t\mathbf{1}_{\{t<\tau\}}+R_\tau\mathbf{1}_{\{\tau\le t\}},\\
Z_t &= z_t\mathbf{1}_{\{t\le\tau\}},\\
U_t &= (R_t-y_t)\mathbf{1}_{\{t\le\tau\}},\\
K_t &= k_{t\wedge\tau}. \tag*{(22)}
\end{aligned} Y t Z t U t K t = y t 1 { t < τ } + R τ 1 { τ ≤ t } , = z t 1 { t ≤ τ } , = ( R t − y t ) 1 { t ≤ τ } , = k t ∧ τ . (22) The random variable R τ R_\tau R τ in this definition is used only on { τ ≤ T } \{\tau\le T\} { τ ≤ T } . Set
Ξ = ξ 1 { τ > T } + R T 1 { τ ≤ T } . (23) \Xi= \xi\mathbf{1}_{\{\tau>T\}}+R_T\mathbf{1}_{\{\tau\le T\}}. \tag*{(23)} Ξ = ξ 1 { τ > T } + R T 1 { τ ≤ T } . (23) Then ( Y , Z , U , K ) (Y,Z,U,K) ( Y , Z , U , K ) satisfies the G \mathbb{G} G -reflected BSDE stopped at T ∧ τ T\wedge\tau T ∧ τ :
Y t ∧ τ = Ξ + ∫ t ∧ τ T ∧ τ f ( s , Y s − , Z s , U s ) d s + K T ∧ τ − K t ∧ τ − ∫ t ∧ τ T ∧ τ Z s d W s − ∫ t ∧ τ T ∧ τ U s d M s . (24) Y_{t\wedge\tau}=\Xi+\int_{t\wedge\tau}^{T\wedge\tau}f(s,Y_{s-},Z_s,U_s)\,\mathrm{d}s+K_{T\wedge\tau}-K_{t\wedge\tau}-\int_{t\wedge\tau}^{T\wedge\tau}Z_s\,\mathrm{d}W_s-\int_{t\wedge\tau}^{T\wedge\tau}U_s\,\mathrm{d}M_s. \tag*{(24)} Y t ∧ τ = Ξ + ∫ t ∧ τ T ∧ τ f ( s , Y s − , Z s , U s ) d s + K T ∧ τ − K t ∧ τ − ∫ t ∧ τ T ∧ τ Z s d W s − ∫ t ∧ τ T ∧ τ U s d M s . (24) Moreover,
Y t ≥ L t on { t < τ } , Y_t \ge L_t\qquad\text{on }\{t<\tau\}, Y t ≥ L t on { t < τ } , and
∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = 0 a.s. \int_{(0,T]} \mathbf{1}_{\{t\le\tau\}}(Y_{t-}-L_{t-})\,\mathrm{d}K_t=0\qquad\text{a.s.} ∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = 0 a.s. The reconstructed processes satisfy the corresponding square-integrability requirements, including
E ∫ 0 T ∧ τ ∣ U t ∣ 2 λ t d t < ∞ . \mathbb{E}\int_0^{T\wedge\tau}|U_t|^2\lambda_t\,\mathrm{d}t<\infty. E ∫ 0 T ∧ τ ∣ U t ∣ 2 λ t d t < ∞. (iii) Converse within the reduction class.
Conversely, suppose that a solution ( Y , Z , U , K ) (Y,Z,U,K) ( Y , Z , U , K ) of (24), satisfying the corresponding obstacle and contact conditions, admits the representation (22) for some
( y , z , k ) ∈ S F 2 × H F 2 × A F 2 , (y,z,k) \in\mathcal{S}^{2}_{F} \times\mathcal{H}^{2}_{F} \times\mathcal{A}^{2}_{F}, ( y , z , k ) ∈ S F 2 × H F 2 × A F 2 , without assuming in advance that ( y , z , k ) (y,z,k) ( y , z , k ) solves the reduced equation. Then ( y , z , k ) (y,z,k) ( y , z , k ) satisfies (19)–(21). Consequently, the reconstructed solution is unique within the specified F \mathbb{F} F -reduction class.
Proof. We divide the proof into four steps.
Step 1: Existence and uniqueness of the reduced equation.
By the definition (18) ,
F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) . F(t,y,z)=f(t,y,z,R_t-y)+\lambda_t(R_t-y). F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) .
The assumptions on f f f , R R R and λ \lambda λ imply that F F F is progressively measurable and uniformly Lipschitz in ( y , z ) (y,z) ( y , z ) .
Indeed, for any ( y , z ) (y,z) ( y , z ) and ( y ′ , z ′ ) (y',z') ( y ′ , z ′ ) ,
∣ F ( t , y , z ) − F ( t , y ′ , z ′ ) ∣ ≤ ∣ f ( t , y , z , R t − y ) − f ( t , y ′ , z ′ , R t − y ′ ) ∣ + λ t ∣ y − y ′ ∣ ≤ C f ( 2 ∣ y − y ′ ∣ + ∣ z − z ′ ∣ ) + λ ˉ ∣ y − y ′ ∣ . \begin{aligned}
\left|F(t,y,z)-F(t,y',z')\right|\leq\left|f(t,y,z,R_t-y)-f(t,y',z',R_t-y')\right|+\lambda_t\left|y-y'\right| \\
&\leq C_f(2|y-y'|+|z-z'|)+\bar{\lambda}|y-y'|.
\end{aligned} ∣ F ( t , y , z ) − F ( t , y ′ , z ′ ) ∣ ≤ ∣ f ( t , y , z , R t − y ) − f ( t , y ′ , z ′ , R t − y ′ ) ∣ + λ t ∣ y − y ′ ∣ ≤ C f ( 2∣ y − y ′ ∣ + ∣ z − z ′ ∣ ) + λ ˉ ∣ y − y ′ ∣.
Thus,
∣ F ( t , y , z ) − F ( t , y ′ , z ′ ) ∣ ≤ C F ( ∣ y − y ′ ∣ + ∣ z − z ′ ∣ ) \left|F(t,y,z)-F(t,y',z')\right|\leq C_F\left(|y-y'|+|z-z'|\right) ∣ F ( t , y , z ) − F ( t , y ′ , z ′ ) ∣ ≤ C F ( ∣ y − y ′ ∣ + ∣ z − z ′ ∣ )
for a deterministic constant C F C_F C F .
Furthermore,
∣ F ( t , 0 , 0 ) ∣ ≤ ∣ f ( t , 0 , 0 , 0 ) ∣ + ( C f + λ ˉ ) ∣ R t ∣ . |F(t,0,0)|\leq|f(t,0,0,0)|+(C_f+\bar{\lambda})|R_t|. ∣ F ( t , 0 , 0 ) ∣ ≤ ∣ f ( t , 0 , 0 , 0 ) ∣ + ( C f + λ ˉ ) ∣ R t ∣.
Consequently,
E ∫ 0 T ∣ F ( t , 0 , 0 ) ∣ 2 d t < ∞ . \mathbb{E}\int_0^T|F(t,0,0)|^2\,dt<\infty. E ∫ 0 T ∣ F ( t , 0 , 0 ) ∣ 2 d t < ∞.
Since L L L is continuous and belongs to S F 2 \mathcal{S}_{\mathbb{F}}^2 S F 2 , ξ ∈ L 2 ( F T ) \xi\in L^2(\mathcal{F}_T) ξ ∈ L 2 ( F T ) , and L T ≤ ξ L_T\leq\xi L T ≤ ξ , the classical existence-and-uniqueness theorem for Brownian reflected BSDEs [13 Theorem 5.2] applies. Therefore, there exists a unique solution ( y , z , k ) (y,z,k) ( y , z , k ) satisfying (19) –(21) .
Step 2: Reconstruction of the enlarged-filtration dynamics.
Introduce
u t = R t − y t . u_t=R_t-y_t. u t = R t − y t .
The reduced BSDE (19) can be written in differential form as
d y t = − [ f ( t , y t , z t , u t ) + λ t u t ] d t − d k t + z t d W t . (25) dy_t=-\left[f(t,y_t,z_t,u_t)+\lambda_tu_t\right]dt-dk_t+z_t\,dW_t. \tag*{(25)} d y t = − [ f ( t , y t , z t , u t ) + λ t u t ] d t − d k t + z t d W t . (25)
Recall that D t = 1 { τ ≤ t } D_t=\mathbf{1}_{\{\tau\leq t\}} D t = 1 { τ ≤ t } . The reconstructed process Y Y Y (first line in (22) ) can be written as
Y t = y t ( 1 − D t ) + ∫ ( 0 , t ] R s d D s . Y_t=y_t(1-D_t)+\int_{(0,t]}R_s\,dD_s. Y t = y t ( 1 − D t ) + ∫ ( 0 , t ] R s d D s .
This representation is valid on [ 0 , T ] [0,T] [ 0 , T ] : the second term is zero before default and equals R τ R_\tau R τ after default.
Applying integration by parts to y t ( 1 − D t ) y_t(1-D_t) y t ( 1 − D t ) yields
d ( y t ( 1 − D t ) ) = ( 1 − D t − ) d y t − y t − d D t − d [ y , D ] t . d\left(y_t(1-D_t)\right)=(1-D_{t-})\,dy_t-y_{t-}\,dD_t-d[y,D]_t. d ( y t ( 1 − D t ) ) = ( 1 − D t − ) d y t − y t − d D t − d [ y , D ] t .
Since y y y is continuous and D D D is of finite variation, we get [ y , D ] t = 0 [y,D]_t=0 [ y , D ] t = 0 . Moreover, y t − = y t y_{t-}=y_t y t − = y t . Therefore,
d ( y t ( 1 − D t ) ) = ( 1 − D t − ) d y t − y t d D t . d\left(y_t(1-D_t)\right)=(1-D_{t-})\,dy_t-y_t\,dD_t. d ( y t ( 1 − D t ) ) = ( 1 − D t − ) d y t − y t d D t .
Adding the differential of ∫ ( 0 , t ] R s d D s \int_{(0,t]}R_s\,dD_s ∫ ( 0 , t ] R s d D s gives
d Y t = ( 1 − D t − ) d y t + ( R t − y t ) d D t . (26) dY_t=(1-D_{t-})\,dy_t+(R_t-y_t)\,dD_t. \tag*{(26)} d Y t = ( 1 − D t − ) d y t + ( R t − y t ) d D t . (26)
Substituting (25) into (26), we obtain
d Y t = − ( 1 − D t − ) f ( t , y t , z t , u t ) d t − ( 1 − D t − ) λ t u t d t − ( 1 − D t − ) d k t + ( 1 − D t − ) z t d W t + u t d D t . dY_t = -(1-D_{t-})f(t,y_t,z_t,u_t)\,dt -(1-D_{t-})\lambda_tu_t\,dt -(1-D_{t-})\,dk_t
+ (1-D_{t-})z_t\,dW_t + u_t\,dD_t. d Y t = − ( 1 − D t − ) f ( t , y t , z t , u t ) d t − ( 1 − D t − ) λ t u t d t − ( 1 − D t − ) d k t + ( 1 − D t − ) z t d W t + u t d D t .
Now, by the compensated-default decomposition (17) , we have d D t = d M t + ( 1 − D t − ) λ t d t dD_t=dM_t+(1-D_{t-})\lambda_t\,dt d D t = d M t + ( 1 − D t − ) λ t d t . Hence u t d D t = u t d M t + ( 1 − D t − ) λ t u t d t u_t\,dD_t=u_t\,dM_t+(1-D_{t-})\lambda_tu_t\,dt u t d D t = u t d M t + ( 1 − D t − ) λ t u t d t . The two terms involving λ t u t d t \lambda_tu_t\,dt λ t u t d t cancel.
Consequently,
d Y t = − ( 1 − D t − ) f ( t , y t , z t , u t ) d t − ( 1 − D t − ) d k t + ( 1 − D t − ) z t d W t + u t d M t . dY_t = -(1-D_{t-})f(t,y_t,z_t,u_t)\,dt -(1-D_{t-})\,dk_t +(1-D_{t-})z_t\,dW_t + u_t\,dM_t. d Y t = − ( 1 − D t − ) f ( t , y t , z t , u t ) d t − ( 1 − D t − ) d k t + ( 1 − D t − ) z t d W t + u t d M t .
Recall that 1 − D t − = 1 { t ≤ τ } 1-D_{t-}=\mathbf{1}_{\{t\le\tau\}} 1 − D t − = 1 { t ≤ τ } for t > 0 t>0 t > 0 . Thus, the reconstructed processes satisfy
U t = u t ( 1 − D t − ) = u t 1 { t ≤ τ } , U_t=u_t(1-D_{t-})=u_t\mathbf{1}_{\{t\le\tau\}}, U t = u t ( 1 − D t − ) = u t 1 { t ≤ τ } ,
Z t = z t ( 1 − D t − ) = z t 1 { t ≤ τ } , Z_t=z_t(1-D_{t-})=z_t\mathbf{1}_{\{t\le\tau\}}, Z t = z t ( 1 − D t − ) = z t 1 { t ≤ τ } ,
K t = k t ∧ τ . K_t=k_{t\wedge\tau}. K t = k t ∧ τ .
Since M M M is constant after τ \tau τ , we have
∫ 0 t u s d M s = ∫ 0 t U s d M s . \int_0^t u_s\,dM_s=\int_0^t U_s\,dM_s. ∫ 0 t u s d M s = ∫ 0 t U s d M s .
Similarly, the Brownian integral is stopped:
∫ 0 t Z s d W s = ∫ 0 t ∧ τ z s d W s . \int_0^t Z_s\,dW_s=\int_0^{t\wedge\tau} z_s\,dW_s. ∫ 0 t Z s d W s = ∫ 0 t ∧ τ z s d W s .
Finally, since k k k is continuous,
d K t = ( 1 − D t − ) d k t , dK_t=(1-D_{t-})\,dk_t, d K t = ( 1 − D t − ) d k t ,
and therefore
K t = ∫ ( 0 , t ] ( 1 − D s − ) d k s = k t ∧ τ . K_t=\int_{(0,t]}(1-D_{s-})\,dk_s=k_{t\wedge\tau}. K t = ∫ ( 0 , t ] ( 1 − D s − ) d k s = k t ∧ τ .
Hence, the three processes retain their pre-default dynamics and are stopped at τ \tau τ , without introducing any additional jump in the reflection process K K K .
Before default,
Y t − = y t , Z t = z t , U t = u t . Y_{t-}=y_t,\qquad Z_t=z_t,\qquad U_t=u_t. Y t − = y t , Z t = z t , U t = u t .
Thus, using the fact that values at t = τ t=\tau t = τ do not affect the Lebesgue integral, we obtain
d Y t = − 1 { t ≤ τ } f ( t , Y t − , Z t , U t ) d t − d K t + Z t d W t + U t d M t . (27) dY_t=-\mathbf{1}_{\{t\le\tau\}}f(t,Y_{t-},Z_t,U_t)\,dt-dK_t+Z_t\,dW_t+U_t\,dM_t. \tag*{(27)} d Y t = − 1 { t ≤ τ } f ( t , Y t − , Z t , U t ) d t − d K t + Z t d W t + U t d M t . (27)
At the default time τ ≤ T \tau\le T τ ≤ T , continuity of y y y and R R R gives
Y τ − = y τ , Y τ = R τ . Y_{\tau-}=y_\tau,\qquad Y_\tau=R_\tau. Y τ − = y τ , Y τ = R τ .
Hence Δ Y τ = R τ − y τ = u τ \Delta Y_\tau=R_\tau-y_\tau=u_\tau Δ Y τ = R τ − y τ = u τ . Since Δ M τ = 1 \Delta M_\tau=1 Δ M τ = 1 , we also have Δ Y τ = U τ Δ M τ \Delta Y_\tau=U_\tau\Delta M_\tau Δ Y τ = U τ Δ M τ . The process K K K is continuous and therefore contributes no jump at default. The reconstructed dynamics are thus valid both before default and at the default time.
Integrating (27) between t ∧ τ t\wedge\tau t ∧ τ and T ∧ τ T\wedge\tau T ∧ τ yields
Y t ∧ τ = Y T ∧ τ + ∫ t ∧ τ T ∧ τ f ( s , Y s − , Z s , U s ) d s + K T ∧ τ − K t ∧ τ − ∫ t ∧ τ T ∧ τ Z s d W s − ∫ t ∧ τ T ∧ τ U s d M s . Y_{t\wedge\tau}=Y_{T\wedge\tau}+\int_{t\wedge\tau}^{T\wedge\tau}f(s,Y_{s-},Z_s,U_s)\,ds+K_{T\wedge\tau}-K_{t\wedge\tau}
-\int_{t\wedge\tau}^{T\wedge\tau}Z_s\,dW_s-\int_{t\wedge\tau}^{T\wedge\tau}U_s\,dM_s. Y t ∧ τ = Y T ∧ τ + ∫ t ∧ τ T ∧ τ f ( s , Y s − , Z s , U s ) d s + K T ∧ τ − K t ∧ τ − ∫ t ∧ τ T ∧ τ Z s d W s − ∫ t ∧ τ T ∧ τ U s d M s .
Finally, if τ > T \tau> T τ > T , then Y T = y T = ξ Y_T = y_T = \xi Y T = y T = ξ , whereas if τ ≤ T \tau\le T τ ≤ T , then Y τ = R τ Y_\tau= R_\tau Y τ = R τ . Therefore, using (23) , we derive
Y T ∧ τ = ξ 1 { τ > T } + R τ 1 { τ ≤ T } = Ξ . Y_{T \wedge\tau} = \xi1_{\{\tau>T\}} + R_\tau1_{\{\tau\le T\}} = \Xi. Y T ∧ τ = ξ 1 { τ > T } + R τ 1 { τ ≤ T } = Ξ.
This proves the enlarged equation (24) .
Step 3: Reflection and square integrability.
Before default, the reconstructed process satisfies Y t = y t Y_t = y_t Y t = y t . Since the reduced solution satisfies y t ≥ L t y_t \ge L_t y t ≥ L t , we obtain Y t ≥ L t Y_t \ge L_t Y t ≥ L t on { t < τ } \{t < \tau\} { t < τ } .
Next, recall that K t = k t ∧ τ K_t = k_{t\wedge\tau} K t = k t ∧ τ . Since k k k is continuous, nondecreasing, and starts from zero, the same properties hold for K K K . Moreover, d K t = 1 { t ≤ τ } d k t dK_t = 1_{\{t\le\tau\}}\,dk_t d K t = 1 { t ≤ τ } d k t . By continuity of y y y and L L L , Y t − − L t − = y t − L t Y_{t-} - L_{t-} = y_t-L_t Y t − − L t − = y t − L t at the times charged by d K dK d K , up to an irrelevant boundary contribution at τ \tau τ . Consequently,
∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = ∫ ( 0 , T ] 1 { t ≤ τ } ( y t − L t ) d k t . \int_{(0,T]} 1_{\{t\le\tau\}}(Y_{t-} - L_{t-})\,dK_t
=
\int_{(0,T]} 1_{\{t\le\tau\}}(y_t - L_t)\,dk_t. ∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = ∫ ( 0 , T ] 1 { t ≤ τ } ( y t − L t ) d k t .
Since the reduced Skorokhod condition gives
∫ ( 0 , T ] ( y t − L t ) d k t = 0 \int_{(0,T]} (y_t - L_t)\,dk_t = 0 ∫ ( 0 , T ] ( y t − L t ) d k t = 0
and the integrand is nonnegative, we conclude that
∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = 0 a.s. \int_{(0,T]} 1_{\{t\le\tau\}}(Y_{t-} - L_{t-})\,dK_t = 0
\qquad\text{a.s.} ∫ ( 0 , T ] 1 { t ≤ τ } ( Y t − − L t − ) d K t = 0 a.s.
We now verify square integrability. Since
∣ Y t ∣ ≤ sup 0 ≤ s ≤ T ∣ y s ∣ + sup 0 ≤ s ≤ T ∣ R s ∣ , |Y_t| \le\sup_{0\le s\le T}|y_s| + \sup_{0\le s\le T}|R_s|, ∣ Y t ∣ ≤ 0 ≤ s ≤ T sup ∣ y s ∣ + 0 ≤ s ≤ T sup ∣ R s ∣ ,
we obtain
E [ sup 0 ≤ t ≤ T ∣ Y t ∣ 2 ] < ∞ . \mathbb{E}\left[\sup_{0\le t\le T}|Y_t|^2\right] < \infty. E [ 0 ≤ t ≤ T sup ∣ Y t ∣ 2 ] < ∞.
Moreover,
E ∫ 0 T ∣ Z t ∣ 2 d t ≤ E ∫ 0 T ∣ z t ∣ 2 d t < ∞ . \mathbb{E}\int_0^T |Z_t|^2\,dt \le\mathbb{E}\int_0^T |z_t|^2\,dt < \infty. E ∫ 0 T ∣ Z t ∣ 2 d t ≤ E ∫ 0 T ∣ z t ∣ 2 d t < ∞.
For the default coefficient,
∣ U t ∣ 2 ≤ 2 ∣ R t ∣ 2 + 2 ∣ y t ∣ 2 . |U_t|^2 \le2|R_t|^2 + 2|y_t|^2. ∣ U t ∣ 2 ≤ 2∣ R t ∣ 2 + 2∣ y t ∣ 2 .
Using the boundedness of λ \lambda λ ,
E ∫ 0 T ∧ τ ∣ U t ∣ 2 λ t d t ≤ 2 λ ˉ T E [ sup 0 ≤ t ≤ T ∣ R t ∣ 2 + sup 0 ≤ t ≤ T ∣ y t ∣ 2 ] < ∞ . \mathbb{E}\int_0^{T\wedge\tau}|U_t|^2\lambda_t\,dt
\le
2\bar{\lambda}T\mathbb{E}\left[\sup_{0\le t\le T}|R_t|^2+\sup_{0\le t\le T}|y_t|^2\right]
<\infty. E ∫ 0 T ∧ τ ∣ U t ∣ 2 λ t d t ≤ 2 λ ˉ T E [ 0 ≤ t ≤ T sup ∣ R t ∣ 2 + 0 ≤ t ≤ T sup ∣ y t ∣ 2 ] < ∞.
Finally,
K T = k T ∧ τ ≤ k T , K_T = k_{T\wedge\tau} \le k_T, K T = k T ∧ τ ≤ k T ,
so
E [ K T 2 ] ≤ E [ k T 2 ] < ∞ . \mathbb{E}[K_T^2] \le\mathbb{E}[k_T^2] < \infty. E [ K T 2 ] ≤ E [ k T 2 ] < ∞.
Thus the reconstructed processes satisfy the required integrability and reflection conditions.
Step 4: Converse and uniqueness within the reduction class.
Suppose conversely that a solution ( Y , Z , U , K ) (Y,Z,U,K) ( Y , Z , U , K ) of (24) admits the reduction form (22) for some
( y , z , k ) ∈ S F 2 × H F 2 × A F 2 . (y,z,k) \in\mathcal{S}_{\mathbb{F}}^2 \times\mathcal{H}_{\mathbb{F}}^2 \times\mathcal{A}_{\mathbb{F}}^2. ( y , z , k ) ∈ S F 2 × H F 2 × A F 2 .
On the event { τ > T } \{\tau> T\} { τ > T } , we have, for every t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] ,
Y t = y t , Z t = z t , U t = R t − y t , K t = k t . Y_t = y_t,\qquad Z_t = z_t,\qquad U_t = R_t-y_t,\qquad K_t = k_t. Y t = y t , Z t = z t , U t = R t − y t , K t = k t .
Moreover, on this event,
M t = − ∫ 0 t λ s d s . M_t=-\int_0^t \lambda_s\,\mathrm{d}s. M t = − ∫ 0 t λ s d s .
Substituting these identities into the enlarged equation yields
y t = ξ + ∫ t T f ( s , y s , z s , R s − y s ) d s + ∫ t T λ s ( R s − y s ) d s + k T − k t − ∫ t T z s d W s . y_t=\xi+\int_t^T f(s,y_s,z_s,R_s-y_s)\,\mathrm{d}s+\int_t^T\lambda_s(R_s-y_s)\,\mathrm{d}s+k_T-k_t-\int_t^T z_s\,\mathrm{d}W_s. y t = ξ + ∫ t T f ( s , y s , z s , R s − y s ) d s + ∫ t T λ s ( R s − y s ) d s + k T − k t − ∫ t T z s d W s .
By the definition of F F F ,
y t = ξ + ∫ t T F ( s , y s , z s ) d s + k T − k t − ∫ t T z s d W s on { τ > T } . y_t=\xi+\int_t^T F(s,y_s,z_s)\,\mathrm{d}s+k_T-k_t-\int_t^T z_s\,\mathrm{d}W_s\quad\text{on }\{\tau>T\}. y t = ξ + ∫ t T F ( s , y s , z s ) d s + k T − k t − ∫ t T z s d W s on { τ > T } .
Define the continuous residual process B B B , with F T \mathcal{F}_T F T -measurable coordinates
B t : = y t − ξ − ∫ t T F ( s , y s , z s ) d s − ( k T − k t ) + ∫ t T z s d W s . B_t:=y_t-\xi-\int_t^T F(s,y_s,z_s)\,\mathrm{d}s-(k_T-k_t)+\int_t^T z_s\,\mathrm{d}W_s. B t := y t − ξ − ∫ t T F ( s , y s , z s ) d s − ( k T − k t ) + ∫ t T z s d W s .
Then B t = 0 B_t=0 B t = 0 for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] on { τ > T } \{\tau>T\} { τ > T } . 1
Let
V = ∣ y T − ξ ∣ + sup t ∈ Q ∩ [ 0 , T ] ∣ B t ∣ , V=\lvert y_T-\xi\rvert+\sup_{t\in\mathbb{Q}\cap[0,T]}\lvert B_t\rvert, V = ∣ y T − ξ ∣ + t ∈ Q ∩ [ 0 , T ] sup ∣ B t ∣ ,
with T T T also included in the supremum if necessary.
By continuity of B B B , the variable V V V detects any failure of the reduced equation. Furthermore, V ≥ 0 V\geq0 V ≥ 0 , V ∈ F T V\in\mathcal{F}_T V ∈ F T , and V 1 { τ > T } = 0 V\mathbf{1}_{\{\tau>T\}}=0 V 1 { τ > T } = 0 a.s.
For every integer n ≥ 1 n\geq1 n ≥ 1 ,
E [ ( V ∧ n ) 1 { τ > T } ] = 0. \mathbb{E}\left[(V\wedge n)\mathbf{1}_{\{\tau>T\}}\right]=0. E [ ( V ∧ n ) 1 { τ > T } ] = 0.
By the Cox survival identity,
P ( τ > T ∣ F T ) = S T = e − Λ T . \mathbb{P}(\tau>T\mid\mathcal{F}_T)=S_T=e^{-\Lambda_T}. P ( τ > T ∣ F T ) = S T = e − Λ T .
Therefore,
0 = E [ ( V ∧ n ) 1 { τ > T } ] = E [ ( V ∧ n ) E [ 1 { τ > T } ∣ F T ] ] = E [ ( V ∧ n ) S T ] . 0=\mathbb{E}\left[(V\wedge n)\mathbf{1}_{\{\tau>T\}}\right]
=\mathbb{E}\left[(V\wedge n)\mathbb{E}\left[\mathbf{1}_{\{\tau>T\}}\mid\mathcal{F}_T\right]\right]
=\mathbb{E}\left[(V\wedge n)S_T\right]. 0 = E [ ( V ∧ n ) 1 { τ > T } ] = E [ ( V ∧ n ) E [ 1 { τ > T } ∣ F T ] ] = E [ ( V ∧ n ) S T ] .
Since S T > 0 S_T>0 S T > 0 a.s., and V ∧ n ≥ 0 V\wedge n\geq0 V ∧ n ≥ 0 , we obtain V ∧ n = 0 V\wedge n=0 V ∧ n = 0 a.s. Letting n → ∞ n\to\infty n → ∞ gives
V = 0 a.s. V=0\qquad\text{a.s.} V = 0 a.s.
Consequently, the reduced equation (19) holds for every t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , almost surely.
We now recover the obstacle condition.
On { τ > T } \{\tau>T\} { τ > T } , the enlarged obstacle condition implies
y t ≥ L t 0 ≤ t ≤ T . y_t\geq L_t\qquad0\leq t\leq T. y t ≥ L t 0 ≤ t ≤ T .
Define
V L = sup 0 ≤ t ≤ T ( L t − y t ) + . V_L=\sup_{0\leq t\leq T}(L_t-y_t)^+. V L = 0 ≤ t ≤ T sup ( L t − y t ) + .
By continuity of L L L and y y y , V L V_L V L is F T \mathcal{F}_T F T -measurable and nonnegative. Moreover,
V L 1 { τ > T } = 0 a.s. V_L \mathbf{1}_{\{\tau>T\}}=0 \quad\text{a.s.} V L 1 { τ > T } = 0 a.s.
Repeating the preceding conditional survival argument yields
V L = 0 a.s. V_L=0 \quad\text{a.s.} V L = 0 a.s.
Hence
y t ≥ L t , 0 ≤ t ≤ T , a.s. y_t \ge L_t,\qquad0\le t\le T,\quad\text{a.s.} y t ≥ L t , 0 ≤ t ≤ T , a.s.
Finally, the enlarged Skorokhod condition implies
∫ ( 0 , T ] 1 { t ≤ τ } ( y t − L t ) d k t = 0 a.s. \int_{(0,T]} \mathbf{1}_{\{t\le\tau\}}(y_t-L_t)\,dk_t=0 \quad\text{a.s.} ∫ ( 0 , T ] 1 { t ≤ τ } ( y t − L t ) d k t = 0 a.s.
By Corollary 4.4 , the Cox survival positivity implies
∫ ( 0 , T ] ( y t − L t ) d k t = 0 a.s. \int_{(0,T]}(y_t-L_t)\,dk_t=0 \quad\text{a.s.} ∫ ( 0 , T ] ( y t − L t ) d k t = 0 a.s.
Thus ( y , z , k ) (y,z,k) ( y , z , k ) satisfies the reduced reflected BSDE together with its obstacle and minimality conditions.
Uniqueness of the reduced Brownian RBSDE implies uniqueness of ( y , z , k ) (y,z,k) ( y , z , k ) . Since the enlarged solution ( Y , Z , U , K ) (Y,Z,U,K) ( Y , Z , U , K ) is determined by (22) , uniqueness follows within the specified reduction class.
This completes the proof.
Remark 5.2 (Meaning and limitations of the converse). The reduction-form hypothesis in the converse is essential.
The theorem does not assert that every solution of the enlarged-filtration RBSDE automatically admits the representation (22). Such a statement would require a separate analysis of martingale representation, reduction of enlarged-filtration processes, and compatibility of the default jump with the prescribed recovery process. Likewise, uniqueness in the full class of enlarged-filtration reflected BSDE solutions is not established by the preceding argument. The theorem proves existence by explicit reconstruction and uniqueness within the specified reference-filtration reduction class.
Our reduction–reconstruction result is closely related to the framework developed by Choulli and Alsheyab [8 ] for optimal stopping under a random horizon.
Their analysis considers a general random time τ \tau τ and its progressive enlargement of F \mathbb{F} F , under the strict positivity of the Azéma supermartingale G G G . In particular, they establish a parametrization of G \mathbb{G} G -optional reward processes in terms of F \mathbb{F} F -observable quantities, derive representations of the corresponding Snell envelopes, and characterize the relationships between optimal stopping times in the two filtrations. The present theorem addresses a related but distinct problem. Rather than starting from the Snell envelope representation, we consider a nonlinear reflected BSDE driven by both a Brownian motion and a compensated default martingale, under a Cox random horizon.
In this setting, the explicit survival process G t = e − Λ t > 0 G_t=e^{-\Lambda_t}>0 G t = e − Λ t > 0 allows us to reconstruct the enlarged-filtration solution from the reduced Brownian reflected BSDE. More precisely, the default jump coefficient is identified by
U t = R t − y t before default , U_t=R_t-y_t \quad\text{before default}, U t = R t − y t before default ,
while the reduced generator takes the form
F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) . F(t,y,z)=f(t,y,z,R_t-y)+\lambda_t(R_t-y). F ( t , y , z ) = f ( t , y , z , R t − y ) + λ t ( R t − y ) .
Furthermore, the survival-support identities established in the preceding sections ensure the transfer of the Skorokhod contact condition, including for general reference Stieltjes reflection measures. Thus, our result complements the optimal stopping approach of Choulli and Alsheyab [8 ] by providing an explicit equation-level reconstruction in the Brownian Cox setting.
Irregular obstacles and Mertens reflection
Reflected BSDEs with irregular optional obstacles may require two distinct increasing processes in the reflection term. In the Mertens formulation, this term typically takes the form
A t + C t − , A_t+C_{t-}, A t + C t − ,
where A A A is predictable and increasing, while C C C is optional, purely discontinuous, and increasing. See, for instance, Grigorova et al. [21 ] .
For each t ∈ ( 0 , T ] t \in(0,T] t ∈ ( 0 , T ] , the left-upper semicontinuous envelope of ζ \zeta ζ is defined by
ζ ‾ t : = lim sup s ↑ t s < t ζ s = lim ε ↓ 0 sup s ∈ [ ( t − ε ) + , t ) ζ s . \overline{\zeta}_t:=\limsup_{\substack{s\uparrow t\\s<t}}\zeta_s=\lim_{\varepsilon\downarrow0}\sup_{s\in[(t-\varepsilon)_+,t)}\zeta_s. ζ t := s ↑ t s < t lim sup ζ s = ε ↓ 0 lim s ∈ [( t − ε ) + , t ) sup ζ s .
In particular, if ζ \zeta ζ admits left limits, then ζ ‾ t = ζ t − \overline{\zeta}_t=\zeta_{t-} ζ t = ζ t − . This envelope allows the predictable contact condition to be formulated even when the obstacle has no left limits.
The two components of A t + C t − A_t+C_{t-} A t + C t − correspond to different contact mechanisms. Depending on the regularity of the obstacle and the chosen formulation, the minimality conditions may involve
H t A = Y t − − ζ ‾ t , H t C = Y t − ζ t . H_t^A=Y_{t-}-\overline{\zeta}_t,\qquad H_t^C=Y_t-\zeta_t. H t A = Y t − − ζ t , H t C = Y t − ζ t .
where ζ ‾ \overline{\zeta} ζ denotes the appropriate left-upper envelope of ζ \zeta ζ .
Irregular-contact transfer
The precise minimality conditions depend on the regularity assumptions and on the decomposition of the predictable reflection component. In the following theorem, we work directly with given nonnegative predictable and optional contact integrands.
Theorem 6.1 (Componentwise irregular-contact transfer). Let A A A and C C C satisfy the standing assumptions on increasing processes introduced in Section 22 . Assume, in addition, that A A A is F \mathbb{F} F -predictable and C C C is F \mathbb{F} F -optional.
Let H A H^A H A be a nonnegative F \mathbb{F} F -predictable process, and let H C H^C H C be a nonnegative F \mathbb{F} F -optional process. Assume that the predictable and optional survival projections introduced in the preceding sections satisfy
q > 0 m A p -a.e. , G > 0 m C o -a.e. q>0\quad m_A^p\text{-a.e.},\qquad G>0\quad m_C^o\text{-a.e.} q > 0 m A p -a.e. , G > 0 m C o -a.e. Then
∫ ( 0 , T ] H t A d A t = 0 a.s. ⟺ ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t = 0 a.s. , \int_{(0,T]}H_t^A\,dA_t=0\quad\text{a.s.}\quad\Longleftrightarrow\quad\int_{(0,T]}H_t^A\mathbf{1}_{\{t\leq\tau\}}\,dA_t=0\quad\text{a.s.}, ∫ ( 0 , T ] H t A d A t = 0 a.s. ⟺ ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t = 0 a.s. , ∫ ( 0 , T ] H t C d C t = 0 a.s. ⟺ ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. \int_{(0,T]}H_t^C\,dC_t=0\quad\text{a.s.}\quad\Longleftrightarrow\quad\int_{(0,T]}H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t C d C t = 0 a.s. ⟺ ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. In particular, both equivalences hold under the Cox construction of Section 4.
Proof. By Theorem 3.3 , the strict positivity of q q q with respect to m A p m_A^p m A p yields
∫ ( 0 , T ] H t A d A t = 0 ⟺ ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t = 0 a.s. \int_{(0,T]} H_t^A\,\mathrm{d}A_t=0 \Longleftrightarrow\int_{(0,T]} H_t^A\mathbf{1}_{\{t\le\tau\}}\,\mathrm{d}A_t=0 \quad\text{a.s.} ∫ ( 0 , T ] H t A d A t = 0 ⟺ ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t = 0 a.s.
Similarly, Theorem 3.1 and the positivity of G G G with respect to m C o m_C^o m C o give
∫ ( 0 , T ] H t C d C t = 0 ⟺ ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. \int_{(0,T]} H_t^C\,\mathrm{d}C_t=0 \Longleftrightarrow\int_{(0,T]} H_t^C\mathbf{1}_{\{t<\tau\}}\,\mathrm{d}C_t=0 \quad\text{a.s.} ∫ ( 0 , T ] H t C d C t = 0 ⟺ ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s.
Under the Cox construction, G t = e − Λ t > 0 G_t=e^{-\Lambda_t}>0 G t = e − Λ t > 0 , and the corresponding predictable survival projection is also strictly positive on the relevant interval. Thus, Corollary 4.4 applies to each reflection component.
Proposition 6.3 (Absence of reflection atoms at a Cox default). Suppose that τ \tau τ is constructed as in § 4 , with continuous cumulative intensity Λ \Lambda Λ . Let C C C be an F \mathbb{F} F -optional, RCLL increasing process satisfying the standing assumptions of § 2 . Then
Δ C τ 1 { 0 < τ ≤ T } = 0 a.s. \Delta C_\tau\mathbf{1}_{\{0<\tau\le T\}}=0 \quad\text{a.s.} Δ C τ 1 { 0 < τ ≤ T } = 0 a.s. Consequently, for every nonnegative F \mathbb{F} F -optional process H H H ,
∫ ( 0 , T ] H t 1 { t < τ } d C t = ∫ ( 0 , T ] H t 1 { t ≤ τ } d C t a.s. \int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,\mathrm{d}C_t=\int_{(0,T]} H_t\mathbf{1}_{\{t\le\tau\}}\,\mathrm{d}C_t \quad\text{a.s.} ∫ ( 0 , T ] H t 1 { t < τ } d C t = ∫ ( 0 , T ] H t 1 { t ≤ τ } d C t a.s.
Proof. Recall that
τ = inf { t ≥ 0 : Λ t ≥ E } , \tau=\inf\{t\ge0:\Lambda_t\ge E\}, τ = inf { t ≥ 0 : Λ t ≥ E } ,
where E E E is a unit exponential random variable independent of F ∞ \mathcal{F}_\infty F ∞ . Since Λ \Lambda Λ is continuous,
{ τ = t } ⊆ { E = Λ t } , t < ∞ . \{\tau=t\}\subseteq\{E=\Lambda_t\},\qquad t<\infty. { τ = t } ⊆ { E = Λ t } , t < ∞.
More generally, the same inclusion holds for any finite F ∞ \mathcal{F}_\infty F ∞ -measurable random time σ \sigma σ . Since Λ σ \Lambda_\sigma Λ σ is F ∞ \mathcal{F}_\infty F ∞ -measurable and E E E is independent of F ∞ \mathcal{F}_\infty F ∞ with an atomless distribution,
P ( τ = σ ) = 0. \mathbb{P}(\tau=\sigma)=0. P ( τ = σ ) = 0.
For each n ≥ 1 n\ge1 n ≥ 1 , the jump times of C C C whose jump sizes exceed 1 / n 1/n 1/ n can be enumerated by a finite or countable family of F \mathbb{F} F -stopping times ( σ n , k ) k ≥ 1 (\sigma_{n,k})_{k\ge1} ( σ n , k ) k ≥ 1 . Therefore,
P ( 0 < τ ≤ T , Δ C τ ≠ 0 ) ≤ ∑ n , k ≥ 1 P ( τ = σ n , k < ∞ ) = 0. \mathbb{P}(0<\tau\le T,\Delta C_\tau\ne0)\le\sum_{n,k\ge1}\mathbb{P}(\tau=\sigma_{n,k}<\infty)=0. P ( 0 < τ ≤ T , Δ C τ = 0 ) ≤ n , k ≥ 1 ∑ P ( τ = σ n , k < ∞ ) = 0.
Finally, the difference between the two Stieltjes integrals is
∫ ( 0 , T ] H t 1 { t ≤ τ } d C t − ∫ ( 0 , T ] H t 1 { t < τ } d C t = H τ Δ C τ 1 { 0 < τ ≤ T } = 0 a.s. \int_{(0,T]} H_t\mathbf{1}_{\{t\le\tau\}}\,\mathrm{d}C_t-\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,\mathrm{d}C_t=H_\tau\Delta C_\tau\mathbf{1}_{\{0<\tau\le T\}}=0 \quad\text{a.s.} ∫ ( 0 , T ] H t 1 { t ≤ τ } d C t − ∫ ( 0 , T ] H t 1 { t < τ } d C t = H τ Δ C τ 1 { 0 < τ ≤ T } = 0 a.s.
This completes the proof.
Corollary 6.4 (Simultaneous transfer of Mertens contact conditions). Under the assumptions of Theorem 6.1, the following equivalence holds:
∫ ( 0 , T ] H t A d A t + ∫ ( 0 , T ] H t C d C t = 0 a.s. \int_{(0,T]} H_t^A\,dA_t+\int_{(0,T]} H_t^C\,dC_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t A d A t + ∫ ( 0 , T ] H t C d C t = 0 a.s. ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t + ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. \int_{(0,T]} H_t^A\mathbf{1}_{\{t\leq\tau\}}\,dA_t+\int_{(0,T]} H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t A 1 { t ≤ τ } d A t + ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. In particular, this equivalence holds under the Cox construction of § 4 .
Proof. All four Stieltjes integrals are nonnegative. Consequently, each sum vanishes almost surely if and only if both of its terms vanish almost surely. The conclusion follows by applying the two equivalences of Theorem 6.1.
Example 6.5 (Failure of contact transfer without survival positivity). Let T = 2 T=2 T = 2 , let F \mathbb{F} F be the trivial filtration, and take the deterministic default time τ = 1 \tau=1 τ = 1 .
Define
C t = 1 { t ≥ 3 / 2 } , H t C = 1. C_t=\mathbf{1}_{\{t\geq3/2\}},\qquad H_t^C=1. C t = 1 { t ≥ 3/2 } , H t C = 1. Then ∫ ( 0 , T ] H t C d C t = 1 \int_{(0,T]}H_t^C\,dC_t=1 ∫ ( 0 , T ] H t C d C t = 1 , whereas ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 \int_{(0,T]}H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t=0 ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 . Thus, ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 \int_{(0,T]}H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t=0 ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 does not imply ∫ ( 0 , T ] H t C d C t = 0 \int_{(0,T]}H_t^C\,dC_t=0 ∫ ( 0 , T ] H t C d C t = 0 . Indeed, the survival process satisfies G t = 1 { t < 1 } G_t=\mathbf{1}_{\{t<1\}} G t = 1 { t < 1 } , so that G = 0 G=0 G = 0 on the support of d C dC d C .
This example illustrates the necessity of the support-positivity condition appearing in Theorem 3.1 and Theorem 6.1.
The preceding results transfer the predictable and optional contact conditions componentwise. In the Cox setting, Proposition 6.3 also shows that the strict and inclusive default-time stopping conventions coincide for the Stieltjes measure d C dC d C .
Pointwise jump contact in Mertens reflection
The componentwise transfer result of Theorem 6.1 is formulated in terms of nonnegative Stieltjes contact integrals. For purely discontinuous reflection, however, the optional minimality condition is often expressed directly at the jump times of the increasing process C C C .
The following proposition shows that these two formulations are equivalent under the optional support condition already established in Theorem 3.1.
Proposition 6.6 (Transfer of pointwise jump-contact minimality). Let C C C satisfy the standing assumptions of Section 2 and suppose, in addition, that C C C is purely discontinuous and F \mathbb{F} F -optional. Let H C H^C H C be a nonnegative F \mathbb{F} F -optional process.
Assume that
G > 0 m C o -a.e. G>0\quad m_C^o\text{-a.e.} G > 0 m C o -a.e. Then the following conditions are equivalent:
H t C Δ C t = 0 for every t ∈ ( 0 , T ] , a.s. H_t^C\Delta C_t=0\quad\text{for every }t\in(0,T],\quad\text{a.s.} H t C Δ C t = 0 for every t ∈ ( 0 , T ] , a.s. and
∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. \int_{(0,T]}H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t=0\quad\text{a.s.} ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. Proof. Since C C C is purely discontinuous and increasing, its Stieltjes measure is concentrated on its jump times. Therefore, for every nonnegative optional process H C H^C H C ,
∫ ( 0 , T ] H t C d C t = ∑ 0 < t ≤ T H t C Δ C t . \int_{(0,T]}H_t^C\,dC_t=\sum_{0<t\leq T}H_t^C\Delta C_t. ∫ ( 0 , T ] H t C d C t = 0 < t ≤ T ∑ H t C Δ C t .
Every term in this sum is nonnegative. Consequently,
∫ ( 0 , T ] H t C d C t = 0 a.s. \int_{(0,T]} H_t^C\,dC_t = 0 \quad\text{a.s.} ∫ ( 0 , T ] H t C d C t = 0 a.s.
if and only if
H t C Δ C t = 0 for every t ∈ ( 0 , T ] , a.s. H_t^C\Delta C_t = 0 \quad\text{for every } t\in(0,T], \quad\text{a.s.} H t C Δ C t = 0 for every t ∈ ( 0 , T ] , a.s.
On the other hand, the positivity assumption
G > 0 m C o -a.e. G>0 \quad m_C^o\text{-a.e.} G > 0 m C o -a.e.
allows us to apply Theorem 3.1 , yielding
∫ ( 0 , T ] H t C d C t = 0 a.s. \int_{(0,T]} H_t^C\,dC_t = 0 \quad\text{a.s.} ∫ ( 0 , T ] H t C d C t = 0 a.s.
if and only if
∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s. \int_{(0,T]} H_t^C\mathbf{1}_{\{t<\tau\}}\,dC_t = 0 \quad\text{a.s.} ∫ ( 0 , T ] H t C 1 { t < τ } d C t = 0 a.s.
Combining these two equivalences completes the proof.
Corollary 6.7. (Pointwise contact for an irregular obstacle .) Suppose that a compatible reduction of an irregular reflected BSDE has already identified the optional contact integrand as
H t C = Y t − ζ t ≥ 0 , H_t^C=Y_t-\zeta_t\geq0, H t C = Y t − ζ t ≥ 0 , where Y Y Y and ζ \zeta ζ are F \mathbb{F} F -optional processes. Under the assumptions of Proposition 6.6 , the following conditions are equivalent:
( Y t − ζ t ) Δ C t = 0 for every t ∈ ( 0 , T ] , a.s. (Y_t-\zeta_t)\Delta C_t=0 \quad\text{for every } t\in(0,T], \quad\text{a.s.} ( Y t − ζ t ) Δ C t = 0 for every t ∈ ( 0 , T ] , a.s. and
∫ ( 0 , T ] ( Y t − ζ t ) 1 { t < τ } d C t = 0 a.s. \int_{(0,T]}(Y_t-\zeta_t)\mathbf{1}_{\{t<\tau\}}\,dC_t=0 \quad\text{a.s.} ∫ ( 0 , T ] ( Y t − ζ t ) 1 { t < τ } d C t = 0 a.s.
Proof. Apply Proposition 6.6 with
H C = Y − ζ . H^C=Y-\zeta. H C = Y − ζ .
Remark 6.8 (Predictable versus optional reflection .) The preceding equivalence concerns only the purely discontinuous optional reflection component C C C . The predictable component A A A is governed by the predictable contact condition involving the appropriate left-upper envelope of the obstacle. Its transfer is already covered by Theorem 3.3 and Theorem 6.1 . Thus, the two reflection components are treated separately, in accordance with their distinct measurability and minimality structures.3
Optional and predictable reflection: distinct minimality structures
The contact-transfer results established above apply to nonnegative integrands of the appropriate measurability class. Their application to reflected BSDEs, however, requires a distinction between optional and predictable obstacles.
For an irregular optional obstacle, the associated strong optional supermartingale may admit a Mertens decomposition involving both a predictable increasing process A A A and a purely discontinuous optional increasing process C C C . The corresponding minimality conditions are governed by different contact mechanisms, as discussed in Theorem 6.1 and Proposition 6.6 .
In contrast, reflected BSDEs with predictable obstacles are formulated through predictable stopping times and require the corresponding predictable notion of reflection. Their minimality conditions cannot, in general, be obtained by replacing an optional obstacle with a predictable one in the usual continuous-obstacle formulation.
Remark 6.9 (Compatibility of contact conventions). The optional and predictable contact-transfer criteria concern the measures and integrands specified in their respective formulations. They do not identify the reflection processes of different reflected BSDE models.
In particular, when a Mertens reflection term is written as A + C − A+C_{-} A + C − , the jump convention for C C C must be distinguished from that of the process C − C_{-} C − .
If C C C is right-continuous, its jumps are Δ C t = C t − C t − \Delta C_t=C_t-C_{t-} Δ C t = C t − C t − , whereas the corresponding right jump of C − C_{-} C − is
Δ + ( C − ) t = C t − C t − = Δ C t . \Delta_{+}(C_{-})_t=C_t-C_{t-}=\Delta C_t. Δ + ( C − ) t = C t − C t − = Δ C t .
Thus, the contact condition must be formulated using the appropriate jump of the increasing process appearing in the chosen Mertens decomposition.
The transfer results of Theorems 3.1 , 3.3 , and 6.1 apply only after these contact integrands and endpoint conventions have been identified.
Quantitative reflection-support transfer and the limits of martingale invariance
The preceding optional and predictable contact-transfer theorems give sharp qualitative criteria for recovering reflection minimality from its restriction to the pre-default interval.
We now derive a quantitative consequence of these criteria. Unlike the equivalence of exact contact identities, this estimate also controls approximate reflection errors and is therefore relevant to approximation and penalisation procedures.
Proposition 7.1 (Quantitative contact transfer). Let ( a t ) (a_t) ( a t ) satisfy the standing assumptions of Section 2, and let H H H be a nonnegative admissible contact integrand.
Consider either of the following settings:
(i) ( a t ) (a_t) ( a t ) and H H H are F \mathbb{F} F -optional, and
G t ≥ ε m a o -a.e. ; G_t \ge\varepsilon m_a^o\text{-a.e.}; G t ≥ ε m a o -a.e. ; (ii) ( a t ) (a_t) ( a t ) and H H H are F \mathbb{F} F -predictable, and
q t ≥ ε m a p -a.e. , q_t \ge\varepsilon m_a^p\text{-a.e.}, q t ≥ ε m a p -a.e. , for some constant ε > 0 \varepsilon>0 ε > 0 .
Then, in the optional case,
ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t 1 { t < τ } d a t ≤ E ∫ ( 0 , T ] H t d a t . \varepsilon\mathbb{E}\int_{(0,T]} H_t\,da_t \le\mathbb{E}\int_{(0,T]} H_t\mathbf{1}_{\{t<\tau\}}\,da_t \le\mathbb{E}\int_{(0,T]} H_t\,da_t. ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t 1 { t < τ } d a t ≤ E ∫ ( 0 , T ] H t d a t .
In the predictable case,
ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t ≤ E ∫ ( 0 , T ] H t d a t . \varepsilon\mathbb{E}\int_{(0,T]} H_t\,\mathrm{d}a_t \le\mathbb{E}\int_{(0,T]} H_t1_{\{t\le\tau\}}\,\mathrm{d}a_t \le\mathbb{E}\int_{(0,T]} H_t\,\mathrm{d}a_t. ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t 1 { t ≤ τ } d a t ≤ E ∫ ( 0 , T ] H t d a t .
The inequalities are understood in the extended nonnegative sense.
Proof. We first consider the optional case.
By the projection identity established in Theorem 3.1 ,
E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t G t d a t . \mathbb{E}\int_{(0,T]} H_t1_{\{t<\tau\}}\,\mathrm{d}a_t=\mathbb{E}\int_{(0,T]}H_tG_t\,\mathrm{d}a_t. E ∫ ( 0 , T ] H t 1 { t < τ } d a t = E ∫ ( 0 , T ] H t G t d a t .
Since
ε ≤ G t ≤ 1 m a o -a.e. , \varepsilon\le G_t\le1\quad m_a^o\text{-a.e.}, ε ≤ G t ≤ 1 m a o -a.e. ,
integration against the nonnegative function H H H gives
ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t G t d a t ≤ E ∫ ( 0 , T ] H t d a t . \varepsilon\mathbb{E}\int_{(0,T]}H_t\,\mathrm{d}a_t\le\mathbb{E}\int_{(0,T]}H_tG_t\,\mathrm{d}a_t\le\mathbb{E}\int_{(0,T]}H_t\,\mathrm{d}a_t. ε E ∫ ( 0 , T ] H t d a t ≤ E ∫ ( 0 , T ] H t G t d a t ≤ E ∫ ( 0 , T ] H t d a t .
This proves the first assertion. The predictable case follows identically from Theorem 3.3 , using the predictable projection q q q in place of G G G .
Corollary 7.2 (Stability of approximate contact conditions). Under either setting of Proposition 7.1, let ( H n ) n ≥ 1 (H^n)_{n\ge1} ( H n ) n ≥ 1 be a sequence of nonnegative contact integrands of the corresponding measurability class.
Suppose that the constant ε > 0 \varepsilon>0 ε > 0 is independent of n n n .
If the corresponding stopped contact errors satisfy
E ∫ ( 0 , T ] H t n 1 { t < τ } d a t ⟶ 0 \mathbb{E}\int_{(0,T]}H_t^n1_{\{t<\tau\}}\,\mathrm{d}a_t\longrightarrow0 E ∫ ( 0 , T ] H t n 1 { t < τ } d a t ⟶ 0 in the optional case, or
E ∫ ( 0 , T ] H t n 1 { t ≤ τ } d a t ⟶ 0 \mathbb{E}\int_{(0,T]}H_t^n1_{\{t\le\tau\}}\,\mathrm{d}a_t\longrightarrow0 E ∫ ( 0 , T ] H t n 1 { t ≤ τ } d a t ⟶ 0 in the predictable case, then
E ∫ ( 0 , T ] H t n d a t ⟶ 0. \mathbb{E}\int_{(0,T]}H_t^n\,\mathrm{d}a_t\longrightarrow0. E ∫ ( 0 , T ] H t n d a t ⟶ 0.
Proof. The conclusion follows immediately from the lower bounds of Proposition 7.1, with the same constant ε \varepsilon ε for every n n n .
the corresponding unstopped integral is strictly positive. Thus, martingale transport alone cannot guarantee recovery of the full reflection minimality condition.
The qualitative support criteria of Theorems 3.1 and 3.3 , and the quantitative bounds established above, remain separate requirements in any general reduction–reconstruction argument.
Example 7.5 (Quantitative transfer for an atomic reflection measure). Consider the Cox construction of Section 4 with constant intensity λ > 0 \lambda> 0 λ > 0 , so that
Λ t = λ t , G t = e − λ t . \Lambda_t = \lambda t,\qquad G_t = e^{-\lambda t}. Λ t = λ t , G t = e − λ t . Fix t 0 ∈ ( 0 , T ) t_0 \in(0,T) t 0 ∈ ( 0 , T ) and let a t = 1 { t ≥ t 0 } a_t = 1_{\{t \ge t_0\}} a t = 1 { t ≥ t 0 } . Thus, d a da d a is concentrated at t 0 t_0 t 0 . For each n ≥ 1 n \ge1 n ≥ 1 , consider the nonnegative F \mathbb{F} F -optional contact error
H t n = 1 n . H_t^n = \frac{1}{n}. H t n = n 1 . The full contact error is
E ∫ ( 0 , T ] H t n d a t = 1 n . \mathbb{E}\int_{(0,T]} H_t^n\,da_t = \frac{1}{n}. E ∫ ( 0 , T ] H t n d a t = n 1 . On the other hand, the stopped contact error satisfies
E ∫ ( 0 , T ] H t n 1 { t < τ } d a t = 1 n P ( τ > t 0 ) = e − λ t 0 n . \mathbb{E}\int_{(0,T]} H_t^n 1_{\{t<\tau\}}\,da_t
= \frac{1}{n}\mathbb{P}(\tau> t_0)
= \frac{e^{-\lambda t_0}}{n}. E ∫ ( 0 , T ] H t n 1 { t < τ } d a t = n 1 P ( τ > t 0 ) = n e − λ t 0 . Consequently,
E ∫ ( 0 , T ] H t n d a t = e λ t 0 E ∫ ( 0 , T ] H t n 1 { t < τ } d a t . \mathbb{E}\int_{(0,T]} H_t^n\,da_t
= e^{\lambda t_0}\mathbb{E}\int_{(0,T]} H_t^n 1_{\{t<\tau\}}\,da_t. E ∫ ( 0 , T ] H t n d a t = e λ t 0 E ∫ ( 0 , T ] H t n 1 { t < τ } d a t . Since t 0 ≤ T t_0 \le T t 0 ≤ T , we also have
E ∫ ( 0 , T ] H t n d a t ≤ e λ T E ∫ ( 0 , T ] H t n 1 { t < τ } d a t ≤ e λ T E ∫ ( 0 , T ] H t n d a t . \mathbb{E}\int_{(0,T]} H_t^n\,da_t
\le e^{\lambda T}\mathbb{E}\int_{(0,T]} H_t^n 1_{\{t<\tau\}}\,da_t
\le e^{\lambda T}\mathbb{E}\int_{(0,T]} H_t^n\,da_t. E ∫ ( 0 , T ] H t n d a t ≤ e λ T E ∫ ( 0 , T ] H t n 1 { t < τ } d a t ≤ e λ T E ∫ ( 0 , T ] H t n d a t . In particular, the stopped contact errors converge to zero if and only if the full contact errors converge to zero.
This example shows that the quantitative support estimate remains valid when the reflection measure is singular with respect to Lebesgue measure and possesses an atom.
Open problem: invariance-time reconstruction beyond Cox models
The Cox reconstruction theorem established above relies on two particularly favorable features: the immersion property and the explicit compensator of the default indicator. These properties allow the martingale, finite-variation, and default-jump terms to be identified separately.
A natural question is whether an analogous reduction–reconstruction principle can be established for the more general invariance times introduced by Crépey and Song [10 ] .
The invariance-time framework. Recall that the framework of Crépey and Song involves two filtrations F ⊆ G \mathbb{F} \subseteq\mathbb{G} F ⊆ G and a G \mathbb{G} G -stopping time τ \tau τ , together with probability measures P \mathbb{P} P and Q \mathbb{Q} Q , where P \mathbb{P} P is equivalent to the restriction of Q \mathbb{Q} Q to F T \mathcal{F}_T F T .
Under the corresponding reduction hypotheses, τ \tau τ is an invariance time if P \mathbb{P} P can be chosen so that every ( F , P ) (\mathbb{F},\mathbb{P}) ( F , P ) -local martingale N N N , stopped immediately before τ \tau τ , becomes a ( G , Q ) (\mathbb{G},\mathbb{Q}) ( G , Q ) -local martingale.
This property transports a class of martingales across both filtration enlargement and change of probability measure. It does not, however, automatically provide a complete representation of the martingale terms appearing in an enlarged-filtration reflected BSDE.
In particular, the invariance property does not by itself identify the default jump, the compensator of the default indicator, or the increasing processes appearing in a Mertens reflection decomposition. These additional structures are essential when the obstacle is irregular and the reflection term contains both predictable and purely discontinuous optional components.
We therefore formulate the following questions, which remain beyond the scope of the preceding results.
Q1. Martingale transport and default compensation. Under what additional structural assumptions on the underlying filtrations can the invariance-time transformation be combined with a martingale representation theorem in the enlarged filtration?4
More precisely, when can the martingale part of an enlarged reflected BSDE be decomposed into a transported reference-filtration martingale and a compensated default martingale, with a well-defined predictable compensator?
Q2. Optional obstacle reduction. Which regularity, measurability, and aggregation conditions ensure that an enlarged-filtration optional obstacle admits a compatible F-optional reduction on the pre-default interval?
In particular, how are its left-upper envelope and the right-jump structure of the corresponding reflected solution preserved under such a reduction?
Q3. Mertens reconstruction. Suppose the enlarged reflected BSDE involves the finite-variation reflection term
A t + C t − , A_t + C_{t-}, A t + C t − ,
where A A A is predictable and increasing, while C C C is optional, purely discontinuous, and increasing.
Under which hypotheses can these two reflection components be reduced and reconstructed separately?
How should possible jumps at τ \tau τ be treated so that the reconstructed reflection term has the correct boundary behavior and creates no spurious mass at the random horizon?
Q4. Existence and uniqueness. Which conditions on the transformed driver, terminal condition, obstacle, and filtration allow the available existence and uniqueness results for reflected BSDEs with irregular obstacles to be applied in the reference filtration?
Under which additional assumptions can the resulting solution be transported back to the enlarged filtration, yielding a genuine equivalence between the two reflected BSDE formulations?
The questions above concern the reconstruction of the entire reflected BSDE, rather than the transfer of its minimality conditions alone.
Theorem 6.1 establishes componentwise transfer of predictable and optional contact identities under the appropriate support conditions. In the Cox setting, the positivity of the survival process yields particularly transparent criteria through Corollary 4.4 .
For a general invariance time, the relevant survival projections, reference measures, and probability measure changes must be handled consistently. Consequently, the contact-transfer results proved in this paper do not, by themselves, imply the existence, uniqueness, or reconstruction of solutions to irregular reflected BSDEs under arbitrary invariance times.
Conclusion
This paper investigated the transfer of reflection minimality conditions between a reference filtration and an enlarged filtration under a random horizon, with particular emphasis on the support of the underlying reflection measures.
We established sharp necessary and sufficient conditions for the equivalence of stopped and unstopped contact identities through optional and predictable dual projections. These criteria identify when stopping at a random time preserves the contact information carried by the reflection measures, including when these measures possess atomic or singular components.
For Cox random times, the explicit conditional survival structure yields a strong transfer principle. We extended the survival-weighting identities to general reference-measurable random measures and derived quantitative estimates under uniform survival positivity. These estimates provide a stability criterion for approximate contact conditions.
In the Brownian Cox framework, we obtained an explicit reconstruction of reflected BSDEs with continuous obstacles. The reduced generator incorporates the default intensity and the prescribed recovery process, while the enlarged solution is reconstructed through its Brownian, default-jump, and reflection components.
For irregular obstacles, we extended the contact-transfer analysis to the predictable and purely discontinuous optional components of Mertens reflection. In particular, we clarified the relationship between integral minimality and pointwise jump-contact conditions, while respecting the distinct measurability and endpoint conventions of the two components.
Scope of the results. The results of this paper establish sharp support criteria for the transfer of optional and predictable reflection-contact conditions, together with their specialization to Cox random times and quantitative stability under uniform survival positivity.
In the Brownian Cox setting, the explicit reconstruction theorem provides existence of an enlarged-filtration solution and uniqueness within the prescribed reference-filtration reduction class. It does not establish uniqueness among all possible enlarged-filtration solutions.
For irregular obstacles, the results concern the transfer of predictable and optional minimality conditions under compatible reductions; they do not establish a general existence or reconstruction theorem.
Extensions to arbitrary invariance times require additional assumptions concerning martingale representation, obstacle reduction, default jumps, and Mertens reflection, as discussed in § 8 .
The main contribution of the paper is therefore to identify precisely which reflection-contact properties can be transferred by measure-theoretic support arguments, and which aspects of a complete reflected BSDE reconstruction require additional stochastic structure.