Introduction
This paper studies how interactive and collective information acquisition about unknown model parameters generates and shapes aggregate learning about them in large-scale systems. Information is dynamically created in a Bayesian framework by costly signals which generate belief processes (posterior, conditional on the signals, distributions). The key feature in our work is that the stochastic dynamics of the signals are directly affected by the collective information-seeking behavior, like duration of learning and assessment about the unknown parameter and decisions when learning terminates. Acquiring information induces a cost consisting of a running component while signals are being used and a final aggregate cost caused by these possibly erroneous assessments and decisions at exit. The focus is on how to (i) minimize the representative information acquisition cost, (ii) investigate through its solution how representative knowledge is being created and shaped by a continuous interaction of signals, (iii) examine its evolution and (iv) build an equilibrium model for its validity. Our results spread across the areas of hypothesis testing, Bayesian learning, filtering, optimal stopping, and mean field games with partial observation.
The motivation to study large-scale models of interactive learning comes from numerous applications in marketing, advertising, trading, autonomous driving systems, and supply-chain systems, to name a few. Indeed, in e-commerce platforms, knowledge about the quality of a new product is generated from reviews and usage data induced by earlier buyers; see Kakhbod, Lanzani and Xing [22 ] . Because some consumers provide detailed feedback while others leave only short comments, the informativeness of the public, collective signals depends considerably on how many and who has chosen to adopt the new product so far. Indeed, if early buyers provide meaningful, detailed feedback, then the signal about the quality of the product is strong. On the other hand, if many early buyers provide only sparse or uninformative feedback, the collective signal is inferior. As a result, the quality of acquired information is being shaped directly by the previous adoption decisions, the population size of this group and the duration of their learning.
In financial markets with liquidity effects, if asset liquidation becomes more intense, the overall market depth deteriorates, making price movements more volatile and less informative about the underlying price fundamentals. As a consequence, the ability to extract information from observed prices depends critically on how many positions are already closed and by when this actually occurred; see, for example, Casgrain and Jaimungal [11 ] . The resulting knowledge acquisition problem is thus affected critically by collective learning interaction, as the value of waiting to gather further information and the quality of information itself are determined by the evolving distribution of decisions and learning duration.
The case of a single decision maker, who learns about a binary random variable Θ \Theta Θ with possible values 0 and 1, has been extensively analyzed in the literature, starting with the work of Wald in [33 ] on costly sequential learning in Bayesian statistics. Since then, it has a long history of development and constitutes a central area in statistical learning and hypothesis testing, with hypotheses being H 0 : Θ = 0 H_0 : \Theta= 0 H 0 : Θ = 0 and H 1 : Θ = 1 H_1 : \Theta= 1 H 1 : Θ = 1 , for this binary case. The underlying problem is to minimize, up to the time that learning stops, the total expected cost consisting of a running part due to signal usage and a terminal component. The latter is generated by erroneous assessments about the values of Θ \Theta Θ at exit time and related decisions. Thus, the controls are formulated in terms of the pair of stopping times and decisions at them, ( τ ( ω ) ; d i ( τ ( ω ) ) ) (\tau(\omega); d_i(\tau(\omega))) ( τ ( ω ) ; d i ( τ ( ω ))) for i = 0 , 1 i = 0, 1 i = 0 , 1 .
Shiryaev developed in [31 ] a connection to an optimal stopping problem for the belief process. In particular, it was shown in [31 ] that the underlying free boundary problem has two continuous and not intersecting boundaries which completely characterize the optimal learning. Specifically, the boundaries divide the space in an information seeking region and two idle regions to its left and right sides. If the belief process starts, at t = 0 t = 0 t = 0 in the left (resp. right) region, no learning occurs and hypothesis H 0 : Θ = 0 H_0 : \Theta= 0 H 0 : Θ = 0 (resp. H 1 : Θ = 1 H_1 : \Theta= 1 H 1 : Θ = 1 ) is instantly accepted. The cost then comes from the instant decision d 0 ( 0 ) d_0(0) d 0 ( 0 ) (resp. d 1 ( 0 ) d_1(0) d 1 ( 0 ) ). If, on the other hand, the belief process starts in the middle region, learning takes place till the first hitting time to either boundary. If the left boundary (resp. right) is reached first at τ ∗ ( ω ) \tau^*(\omega) τ ∗ ( ω ) , hypothesis H 0 : Θ = 0 H_0 : \Theta= 0 H 0 : Θ = 0 (resp. H 1 : Θ = 1 H_1 : \Theta= 1 H 1 : Θ = 1 ) is accepted and decision d 0 ( τ ∗ ( ω ) ) d_0(\tau^*(\omega)) d 0 ( τ ∗ ( ω )) (resp. d 1 ( τ ∗ ( ω ) ) d_1(\tau^*(\omega)) d 1 ( τ ∗ ( ω )) is made. The total cost is the one incurred by the latter and the running learning cost in [ 0 , τ ∗ ( ω ) ] [0,\tau^*(\omega)] [ 0 , τ ∗ ( ω )] .
The results of [31 ] initiated a very rich and active line of research, and providing a complete bibliography is beyond the scope herein (see, among others, Moscarini and Smith [28 ] , Miao and Xing [27 ] , Ke, Tang, Villas-Boas and Zhang [24 ] ). We also mention the works of Ekström, Karatzas and Vaicenavicius [17 ] , Ekström and Vaicenavicius [18 ] , Ekström and Wang [19 ] and Campbell, Gaitsgori, Groenewald and Karatzas [7 ] .
Herein, we build a mean field game of interactive and collective information acquisition, which we study using the technical approach of [31 ] . We assume that information is being acquired and propagated through a continuous interaction within a population continuum which uses signals with dynamics directly affected by the aggregate learning, decisions and costs. This continuum starts with a distribution of prior probabilities for the values of Θ \Theta Θ and interacts through time till learning stops. The key quantity is the temporal aggregate of the fraction of the population that has completed learning, selected a decision, encountered the relevant costs and exited. This temporal aggregate alters the learning mechanism of the remaining population through the drift of its, continuously updated, signal. In turn, it affects the representative value function which models the minimal total expected informational cost and constitutes the best response to the effects of realized learning and costs. We are then motivated to introduce a fixed-point notion in terms of this aggregate feedback which leads to a mean field game.
We stress that contrary to the standard mean field games where players are entirely identified by their controlled state processes, herein the notion of a player is identified with the individual prior probability and subsequent belief process. In other words, there is no representative controlled diffusion as such but rather a representative belief process whose dynamics are affected by the learning and decisions of the peer continuum. To our knowledge, the only work close to ours is the preprint of Campbell and Zhang [8 ] , but both their technical assumptions and modeling elements are different than ours. We provide further details at the end of the introduction.
The model and the results. In what follows we explain our main results and refer for the precise definitions and assumptions later in the paper.
Let ( Ω , F , P ) (\Omega,\mathcal{F},\mathbb{P}) ( Ω , F , P ) be a probability space that supports
(i) a binary random variable Θ ∈ { 0 , 1 } \Theta\in\{0,1\} Θ ∈ { 0 , 1 } ,
(ii) a random variable P 0 ∈ [ 0 , 1 ] P_0 \in[0,1] P 0 ∈ [ 0 , 1 ] with law η \eta η having a continuous density representing the initial belief (prior) that P ( Θ = 1 ∣ P 0 ) = P 0 \mathbb{P}(\Theta=1 \mid P_0)=P_0 P ( Θ = 1 ∣ P 0 ) = P 0 ,
(iii) a standard Brownian motion W W W .
(i) a binary random variable Θ ∈ { 0 , 1 } , (ii) a random variable P 0 ∈ [ 0 , 1 ] with law η having a continuous density representing the initial belief (prior) that P ( Θ = 1 ∣ P 0 ) = P 0 , (iii) a standard Brownian motion W . (1) \begin{aligned}
&\text{(i) a binary random variable }\Theta\in\{0,1\},\\
&\text{(ii) a random variable }P_0 \in[0,1]\text{ with law }\eta\text{ having a continuous density}\\
&\qquad\text{representing the initial belief (prior) that }\mathbb{P}(\Theta=1\mid P_0)=P_0,\\
&\text{(iii) a standard Brownian motion }W. \tag*{(1)}
\end{aligned} (i) a binary random variable Θ ∈ { 0 , 1 } , (ii) a random variable P 0 ∈ [ 0 , 1 ] with law η having a continuous density representing the initial belief (prior) that P ( Θ = 1 ∣ P 0 ) = P 0 , (iii) a standard Brownian motion W . (1)
It is assumed that
W is independent of ( Θ , P 0 ) . (2) W\text{ is independent of }(\Theta,P_0). \tag*{(2)} W is independent of ( Θ , P 0 ) . (2)
We study information acquisition about the unknown binary state Θ \Theta Θ from continuous time signals in a large-scale system. Knowledge is being created from a representative observational signal Y μ Y^\mu Y μ whose dynamics depend on aggregate learning interaction in the continuum. Specifically, Y μ Y^\mu Y μ is taken to solve
d Y t μ = b ( t , Θ ; F ( t ; μ ≤ t ) ) d t + d W t for t ∈ [ 0 , T ] and Y 0 μ = 0. (3) \mathrm{d}Y_t^\mu=b(t,\Theta;F(t;\mu_{\leq t}))\,\mathrm{d}t+\mathrm{d}W_t\quad\text{for }t\in[0,T]\text{ and }Y_0^\mu=0. \tag*{(3)} d Y t μ = b ( t , Θ ; F ( t ; μ ≤ t )) d t + d W t for t ∈ [ 0 , T ] and Y 0 μ = 0. (3)
The term F ( t ; μ ≤ t ) F(t;\mu_{\leq t}) F ( t ; μ ≤ t ) , which is defined in (16) , models the collective information feedback and incorporates the learning duration, exit times, and decisions about the unknown values and related costs up to time t t t . The input μ \mu μ is a pair ( μ 0 , μ 1 ) (\mu^0,\mu^1) ( μ 0 , μ 1 ) of two nonnegative subadditive measures representing stopping times and decisions at them, given, for each Borel subset A A A of [ 0 , T ] [0,T] [ 0 , T ] and i = 0 , 1 i=0,1 i = 0 , 1 , by
μ i ( A ) = P ( τ ∈ A , d = i ) , \mu^i(A)=\mathbb{P}(\tau\in A,d=i), μ i ( A ) = P ( τ ∈ A , d = i ) ,
and such that μ 0 + μ 1 \mu^0+\mu^1 μ 0 + μ 1 is a probability measure on [ 0 , T ] [0,T] [ 0 , T ] . In what follows, we denote the set of such pairs by M \mathcal{M} M , that is,
M : = { ( μ 0 , μ 1 ) : μ 0 , μ 1 ∈ M + ( [ 0 , T ] ) , μ 0 + μ 1 ∈ P ( [ 0 , T ] ) } , \mathcal{M}:=\left\{(\mu^0,\mu^1):\mu^0,\mu^1\in\mathcal{M}_+([0,T]),\ \mu^0+\mu^1\in\mathcal{P}([0,T])\right\}, M := { ( μ 0 , μ 1 ) : μ 0 , μ 1 ∈ M + ([ 0 , T ]) , μ 0 + μ 1 ∈ P ([ 0 , T ]) } ,
where M + ( [ 0 , T ] ) \mathcal{M}_{+}([0,T]) M + ([ 0 , T ]) stands for the set of nonnegative subadditive measures on [ 0 , T ] [0,T] [ 0 , T ] .
For a fixed input μ \mu μ , the observation signal Y μ Y^\mu Y μ generates the representative posterior belief process P P P given by
P t : = P ( Θ = 1 ∣ F t μ ) for t ∈ [ 0 , T ] and P 0 = p ∈ [ 0 , 1 ] , P_t := \mathbb{P}(\Theta= 1 \mid\mathcal{F}^{\mu}_t) \text{ for } t \in[0,T] \text{ and } P_0 = p \in[0,1], P t := P ( Θ = 1 ∣ F t μ ) for t ∈ [ 0 , T ] and P 0 = p ∈ [ 0 , 1 ] ,
where F μ = ( F t μ ) 0 ≤ t ≤ T \mathcal{F}^{\mu} = (\mathcal{F}^{\mu}_t)_{0\leq t\leq T} F μ = ( F t μ ) 0 ≤ t ≤ T denotes the usual augmentation of the filtration ( σ ( P 0 , Y s μ ) : 0 ≤ s ≤ t ) ) 0 ≤ t ≤ T (\sigma(P_0,Y^\mu_s):0\leq s\leq t))_{0\leq t\leq T} ( σ ( P 0 , Y s μ ) : 0 ≤ s ≤ t ) ) 0 ≤ t ≤ T .
Classical results in filtering yield that the process P P P solves the sde
d P t μ = Δ b μ ( t ) P t μ ( 1 − P t μ ) d Z t μ for t ∈ ( 0 , T ] and P 0 = p ∈ [ 0 , 1 ] , (4) \mathrm{d}P^\mu_t = \Delta b^\mu(t)P^\mu_t(1-P^\mu_t)\,\mathrm{d}Z^\mu_t \quad\text{for } t \in(0,T] \text{ and } P_0 = p \in[0,1], \tag*{(4)} d P t μ = Δ b μ ( t ) P t μ ( 1 − P t μ ) d Z t μ for t ∈ ( 0 , T ] and P 0 = p ∈ [ 0 , 1 ] , (4)
where
Δ b μ ( t ) : = Δ b ( t ; F ( t ; μ ≤ t ) ) = b ( t , 1 ; F ( t ; μ ≤ t ) ) − b ( t , 0 ; F ( t ; μ ≤ t ) ) (5) \Delta b^\mu(t) := \Delta b(t;F(t;\mu_{\leq t})) = b(t,1;F(t;\mu_{\leq t})) - b(t,0;F(t;\mu_{\leq t})) \tag*{(5)} Δ b μ ( t ) := Δ b ( t ; F ( t ; μ ≤ t )) = b ( t , 1 ; F ( t ; μ ≤ t )) − b ( t , 0 ; F ( t ; μ ≤ t )) (5)
is the difference between the signal drifts under the two states 0 and 1, and Z μ Z^\mu Z μ is a F μ \mathcal{F}^\mu F μ -Brownian motion referred to as the innovation process; see, for example, Bain and Crisan [1 ] .
We assume (see (17) and (18) ) that b b b and Δ b \Delta b Δ b satisfy the assumptions that are needed for (3) and (4) to have unique strong solutions.
The representative information acquisition problem is an optimization one over all stopping learning times and decisions related to parameter value assessment.
Following [31 ] , we formulate the original problem as an optimal stopping problem for the representative belief process over all stopping learning times.
Specifically, we introduce the value function
V μ ( 0 , p ) = inf τ ∈ T 0 , T μ E [ ∫ 0 τ f μ ( s , P s μ ) d s + G ( P τ μ ) : P 0 μ = p ] (6) V^\mu(0,p) = \inf_{\tau\in\mathcal{T}^\mu_{0,T}} \mathbb{E}\left[\int_0^\tau f^\mu(s,P^\mu_s)\,\mathrm{d}s + G(P^\mu_\tau):P^\mu_0=p\right] \tag*{(6)} V μ ( 0 , p ) = τ ∈ T 0 , T μ inf E [ ∫ 0 τ f μ ( s , P s μ ) d s + G ( P τ μ ) : P 0 μ = p ] (6)
where, for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , T t , T μ \mathcal{T}^\mu_{t,T} T t , T μ denotes the set of [ t , T ] [t,T] [ t , T ] -valued F μ \mathbb{F}^\mu F μ -stopping times, f μ : [ 0 , T ] × [ 0 , 1 ] → R f^\mu:[0,T]\times[0,1]\to\mathbb{R} f μ : [ 0 , T ] × [ 0 , 1 ] → R is the running cost during learning given by
f μ ( t , p ) : = f ( t , p ; F ( t ; μ ≤ t ) ) , (7) f^\mu(t,p):=f(t,p;F(t;\mu_{\leq t})), \tag*{(7)} f μ ( t , p ) := f ( t , p ; F ( t ; μ ≤ t )) , (7)
with f f f satisfying (20) , and G : [ 0 , 1 ] → R G:[0,1]\to\mathbb{R} G : [ 0 , 1 ] → R is the penalty incurred when learning stops, taken to be
G ( p ) : = min i = 0 , 1 G ‾ ( p , i ) . G(p):=\min_{i=0,1}\overline{G}(p,i). G ( p ) := i = 0 , 1 min G ( p , i ) .
Following [31 ] and subsequent papers, we assume that, for a 0 , a 1 > 0 a_0,a_1>0 a 0 , a 1 > 0 ,
G ‾ ( p , 0 ) = a 0 p and G ‾ ( p , 1 ) = a 1 ( 1 − p ) . \overline{G}(p,0)=a_0p \quad\text{and} \quad\overline{G}(p,1)=a_1(1-p). G ( p , 0 ) = a 0 p and G ( p , 1 ) = a 1 ( 1 − p ) .
With this choice of G ‾ \overline{G} G , the terminal cost G G G can be written as
G ( p ) = min { a 0 p , a 1 ( 1 − p ) } . (8) G(p)=\min\{a_0p,a_1(1-p)\}. \tag*{(8)} G ( p ) = min { a 0 p , a 1 ( 1 − p )} . (8)
The maximum in (8) is attained at
m = a 1 a 0 + a 1 . (9) m=\frac{a_1}{a_0+a_1}. \tag*{(9)} m = a 0 + a 1 a 1 . (9)
Then, we can rewrite G G G as
G ( p ) = G ‾ ( p , D ( p ) ) , G(p)=\overline{G}(p,D(p)), G ( p ) = G ( p , D ( p )) ,
where D : [ 0 , 1 ] → { 0 , 1 } D:[0,1]\to\{0,1\} D : [ 0 , 1 ] → { 0 , 1 } is the so-called decision rule, with the tie at m m m assigned to decision 0,
D ( p ) : = { 0 if p ≤ m , 1 if p > m . D(p):=
\begin{cases}
0 & \text{if }p\le m,\\
1 & \text{if }p>m.
\end{cases} D ( p ) := { 0 1 if p ≤ m , if p > m .
We remark that it is possible to work with other choices of G ‾ ( ⋅ , 0 ) \overline{G}(\cdot,0) G ( ⋅ , 0 ) and G ‾ ( ⋅ , 1 ) \overline{G}(\cdot,1) G ( ⋅ , 1 ) as long as each one of them is concave and G ‾ ( 0 , 0 ) < G ‾ ( 0 , 1 ) \overline{G}(0,0)<\overline{G}(0,1) G ( 0 , 0 ) < G ( 0 , 1 ) and G ‾ ( 1 , 0 ) > G ‾ ( 1 , 1 ) \overline{G}(1,0)>\overline{G}(1,1) G ( 1 , 0 ) > G ( 1 , 1 ) , provided we make assumptions on the sign of ∂ p 2 G \partial_p^2G ∂ p 2 G near the maximum. In addition, we can assume that the constants a 0 , a 1 a_0,a_1 a 0 , a 1 may depend on time. Here, for the sake of simplicity and in order not to confuse the reader with unnecessary mathematical technicalities, we have chosen to work with the classical choice of G G G in (8) .
The population learning feedback μ ∈ M \mu\in\mathcal{M} μ ∈ M influences not only the dynamics of P μ P^\mu P μ but, also, the running cost, and V μ V^\mu V μ is interpreted as a best response towards μ \mu μ of the knowledge creation problem. This leads to a notion of mean field equilibrium, which we formulate in the sequel.
For each fixed μ ∈ M \mu\in\mathcal{M} μ ∈ M , the time dependent value function V μ : [ 0 , T ] × [ 0 , 1 ] → R V^\mu:[0,T]\times[0,1]\to\mathbb{R} V μ : [ 0 , T ] × [ 0 , 1 ] → R is defined, in complete analogy to (6) , by
V μ ( t , p ) = inf τ ∈ T t , T μ E [ ∫ t τ f μ ( s , P s t , p ; μ ) d s + G ( P τ t , p ; μ ) ] . V^\mu(t,p)=\inf_{\tau\in\mathcal{T}_{t,T}^\mu}\mathbb{E}\left[\int_t^\tau f^\mu(s,P_s^{t,p;\mu})\,\mathrm{d}s+G(P_\tau^{t,p;\mu})\right]. V μ ( t , p ) = τ ∈ T t , T μ inf E [ ∫ t τ f μ ( s , P s t , p ; μ ) d s + G ( P τ t , p ; μ ) ] .
where P t , p ; μ P^{t,p;\mu} P t , p ; μ solves (4) in [ t , T ] [t,T] [ t , T ] and P t t , p ; μ = p P_t^{t,p;\mu}=p P t t , p ; μ = p .
We establish that V μ V^\mu V μ is the unique (viscosity) solution to the terminal value variational inequality
max { − ∂ t v − a μ ∂ p 2 v − f μ , v − G } = 0 in [ 0 , T ) × [ 0 , 1 ] v ( T , p ) = G ( p ) , (10) \max\left\{-\partial_t v-a^\mu\partial_p^2v-f^\mu,v-G\right\}=0\quad\text{in }[0,T)\times[0,1]\qquad v(T,p)=G(p), \tag*{(10)} max { − ∂ t v − a μ ∂ p 2 v − f μ , v − G } = 0 in [ 0 , T ) × [ 0 , 1 ] v ( T , p ) = G ( p ) , (10)
where
a μ ( t , p ) = 1 2 ( Δ b μ ( t ) ) 2 p 2 ( 1 − p ) 2 . (11) a^\mu(t,p)=\frac{1}{2}\left(\Delta b^\mu(t)\right)^2p^2(1-p)^2. \tag*{(11)} a μ ( t , p ) = 2 1 ( Δ b μ ( t ) ) 2 p 2 ( 1 − p ) 2 . (11)
We show that, for each μ ∈ M \mu\in\mathcal{M} μ ∈ M , the domain [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] is divided in two regions, namely, the exploration region
E μ = { ( t , p ) ∈ [ 0 , T ] × [ 0 , 1 ] : V ( t , p ) < G ( p ) } , (12) \mathcal{E}^\mu=\{(t,p)\in[0,T]\times[0,1]:V(t,p)<G(p)\}, \tag*{(12)} E μ = {( t , p ) ∈ [ 0 , T ] × [ 0 , 1 ] : V ( t , p ) < G ( p )} , (12)
and the stopping region
S μ = { ( t , p ) ∈ [ 0 , T ] × [ 0 , 1 ] : V ( t , p ) = G ( p ) } . (13) \mathcal{S}^\mu=\{(t,p)\in[0,T]\times[0,1]:V(t,p)=G(p)\}. \tag*{(13)} S μ = {( t , p ) ∈ [ 0 , T ] × [ 0 , 1 ] : V ( t , p ) = G ( p )} . (13)
In what follows, for each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , we write E t μ \mathcal{E}_t^\mu E t μ and S t μ \mathcal{S}_t^\mu S t μ for the t t t –slices of the exploration and stopping regions respectively, and note that E T μ = ∅ \mathcal{E}_T^\mu=\varnothing E T μ = ∅ and S T μ = [ 0 , 1 ] \mathcal{S}_T^\mu=[0,1] S T μ = [ 0 , 1 ] . We show that
E t μ = ( A − μ ( t ) , A + μ ( t ) ) and S t μ = [ 0 , A − μ ( t ) ] ∪ [ A + μ ( t ) , 1 ] , \mathcal{E}_t^\mu=(A_-^\mu(t),A_+^\mu(t))\quad\text{and}\quad\mathcal{S}_t^\mu=[0,A_-^\mu(t)]\cup[A_+^\mu(t),1], E t μ = ( A − μ ( t ) , A + μ ( t )) and S t μ = [ 0 , A − μ ( t )] ∪ [ A + μ ( t ) , 1 ] ,
where, for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
A − μ ( t ) : = sup { p ∈ S t μ ∩ [ 0 , m ] } and A + μ ( t ) : = inf { p ∈ S t μ ∩ ( m , 1 ] } . (14) A_-^\mu(t):=\sup\{p\in\mathcal{S}_t^\mu\cap[0,m]\}\quad\text{and}\quad A_+^\mu(t):=\inf\{p\in\mathcal{S}_t^\mu\cap(m,1]\}. \tag*{(14)} A − μ ( t ) := sup { p ∈ S t μ ∩ [ 0 , m ]} and A + μ ( t ) := inf { p ∈ S t μ ∩ ( m , 1 ]} . (14)
The maps t → A − μ ( t ) t\to A_-^\mu(t) t → A − μ ( t ) and t → A + μ ( t ) t\to A_+^\mu(t) t → A + μ ( t ) are, respectively, the left and right free boundaries of V μ V^\mu V μ . We show that the value function is uniformly stable with respect to the weak convergence of M \mathcal{M} M , that is, if μ n , μ ∈ M \mu^{n},\mu\in\mathcal{M} μ n , μ ∈ M are such that μ n ⇀ μ \mu^{n}\rightharpoonup\mu μ n ⇀ μ , then the value functions V μ n V^{\mu^{n}} V μ n converge uniformly on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] to V μ V^{\mu} V μ .
In turn, we combine this continuity with various local regularity and non-degeneracy estimates to prove that there exists R ⊂ ( 0 , T ) \mathcal{R}\subset(0,T) R ⊂ ( 0 , T ) , consisting of the times where the interfaces [ 0 , T ] ∋ t ↦ A ± μ ( t ) [0,T]\ni t\mapsto A^{\mu}_{\pm}(t) [ 0 , T ] ∋ t ↦ A ± μ ( t ) are continuous, of full measure such that, as μ n ⇀ μ \mu^{n}\rightharpoonup\mu μ n ⇀ μ and locally uniformly in R \mathcal{R} R ,
A ± μ n ⟶ A ± μ . A^{\mu^{n}}_{\pm}\longrightarrow A^{\mu}_{\pm}. A ± μ n ⟶ A ± μ .
In the sequel, we use the best response pair ( τ ^ 0 , p μ , d ^ 0 , p μ ) (\hat{\tau}^{\mu}_{0,p},\hat{d}^{\mu}_{0,p}) ( τ ^ 0 , p μ , d ^ 0 , p μ ) , which, for each initial belief p p p and information input μ ∈ M \mu\in\mathcal{M} μ ∈ M , is given by
τ ^ 0 , p μ : = inf { t ∈ [ 0 , T ] : V μ ( t , P t 0 , p ) = G ( P t 0 , p ) } and d ^ 0 , p μ : = D ( P τ ^ 0 , p μ 0 , p ) . (15) \hat{\tau}^{\mu}_{0,p}:=\inf\{t\in[0,T]:V^{\mu}(t,P^{0,p}_{t})=G(P^{0,p}_{t})\}\quad\text{and}\quad\hat{d}^{\mu}_{0,p}:=D(P^{0,p}_{\hat{\tau}^{\mu}_{0,p}}). \tag*{(15)} τ ^ 0 , p μ := inf { t ∈ [ 0 , T ] : V μ ( t , P t 0 , p ) = G ( P t 0 , p )} and d ^ 0 , p μ := D ( P τ ^ 0 , p μ 0 , p ) . (15)
For each Borel set A ⊂ [ 0 , T ] A\subset[0,T] A ⊂ [ 0 , T ] and i = 0 , 1 i=0,1 i = 0 , 1 , we define
ν i μ ( A , p ) : = P ( τ ^ 0 , p μ ∈ A , d ^ 0 , p μ = i ) and ν i μ ( A ) : = ∫ [ 0 , 1 ] ν i μ ( A , p ) d η ( p ) , \nu^{\mu}_{i}(A,p):=\mathbb{P}(\hat{\tau}^{\mu}_{0,p}\in A,\ \hat{d}^{\mu}_{0,p}=i)\quad\text{and}\quad\nu^{\mu}_{i}(A):=\int_{[0,1]}\nu^{\mu}_{i}(A,p)\,\mathrm{d}\eta(p), ν i μ ( A , p ) := P ( τ ^ 0 , p μ ∈ A , d ^ 0 , p μ = i ) and ν i μ ( A ) := ∫ [ 0 , 1 ] ν i μ ( A , p ) d η ( p ) ,
and note that ( ν 0 , ν 1 ) ∈ M ( 0 , T ) (\nu_{0},\nu_{1})\in\mathcal{M}(0,T) ( ν 0 , ν 1 ) ∈ M ( 0 , T ) .
Next, we introduce the map
M ∋ μ ↦ Φ ( μ ) : = ( ν 0 μ , ν 1 μ ) ∈ M , \mathcal{M}\ni\mu\mapsto\Phi(\mu):=(\nu^{\mu}_{0},\nu^{\mu}_{1})\in\mathcal{M}, M ∋ μ ↦ Φ ( μ ) := ( ν 0 μ , ν 1 μ ) ∈ M ,
and define a mean field (solution) equilibrium to be a fixed point
Φ ( μ ∗ ) = μ ∗ . \Phi(\mu^{*})=\mu^{*}. Φ ( μ ∗ ) = μ ∗ .
Equivalently, for i = 0 , 1 i=0,1 i = 0 , 1 and any Borel A ⊂ [ 0 , T ] A\subset[0,T] A ⊂ [ 0 , T ] , the components ( μ 0 ∗ ( A ) , μ 1 ∗ ( A ) ) (\mu^{*}_{0}(A),\mu^{*}_{1}(A)) ( μ 0 ∗ ( A ) , μ 1 ∗ ( A )) of μ ∗ \mu^{*} μ ∗ satisfy the consistency relationship
μ i ∗ ( A ) = ∫ [ 0 , 1 ] P ( τ ^ 0 , p μ ∗ ∈ A , d ^ 0 , p μ ∗ = i ) d η ( p ) . \mu^{*}_{i}(A)=\int_{[0,1]}\mathbb{P}(\hat{\tau}^{\mu^{*}}_{0,p}\in A,\ \hat{d}^{\mu^{*}}_{0,p}=i)\,\mathrm{d}\eta(p). μ i ∗ ( A ) = ∫ [ 0 , 1 ] P ( τ ^ 0 , p μ ∗ ∈ A , d ^ 0 , p μ ∗ = i ) d η ( p ) .
The main result of the paper is stated next.
Theorem 1.1. Assume (22). Then, the map M ∋ μ ↦ Φ ( μ ) \mathcal{M}\ni\mu\mapsto\Phi(\mu) M ∋ μ ↦ Φ ( μ ) has a fixed point.
For each candidate population learning distribution μ ∈ M \mu\in\mathcal{M} μ ∈ M , the joint law of the best response defines Φ ( μ ) ∈ M \Phi(\mu)\in\mathcal{M} Φ ( μ ) ∈ M . We show that the map Φ : M → M \Phi:\mathcal{M}\to\mathcal{M} Φ : M → M admits a fixed point μ ∗ \mu^{*} μ ∗ , which represents a mean field game equilibrium.
We conclude showing that the mean field model provides a quantitative approximation to its finite population analogue. Among others, we show that the mean field population learning feedback is an O ( ( N − 1 ) − 1 / 2 ) O((N-1)^{-1/2}) O (( N − 1 ) − 1/2 ) approximate Nash equilibrium for the population of size N N N .
Related literature. There has been a substantial body of work on mean field games with optimal stopping in the fully observable setting. Among others, In [29 ] , Nutz studied interactions through the number of agents who have already stopped, while Carmona, Delarue and Lacker in [10 ] analyzed a bank-run game with Brownian common noise and proved existence of mean field equilibria. The first general analytical treatment using the coupled Hamilton–Jacobi–Bellman and Fokker–Planck system was given by Bertucci in [2 ] , who introduced the notion of mixed solutions. A “dual” linear programming approach involving relaxed solutions was developed by Bouveret, Dumitrescu and Tankov in [6 ] and Dumitrescu, Leutscher and Tankov [15 , 16 ] . Further advances include a master equation formulation by Possamaï and Talbi [30 ] , a mean field Bank–El Karoui representation approach by He, Tan and Zou [21 ] , and related work on McKean–Vlasov optimal stopping by Talbi, Touzi and Zhang [32 ] , Cosso and Perelli [12 ] , and Cardaliaguet, Jackson and Souganidis [9 ] .
The closest work to our formulation is [8 ] . Their main analysis concerns the objective E [ c τ + G ( P τ ) ] \mathbb{E}[c\tau+G(P_{\tau})] E [ c τ + G ( P τ )] and requires the obstacle G G G to be smooth, concave, and symmetric about 1 / 2 1/2 1/2 . The classical piecewise-affine obstacle is treated separately but only under an additional monotonicity assumption on the drift of the belief process. Their proof of the boundary stability directly adapts the probabilistic approach of De Angelis and Stabile [14 ] , which necessitates additional structural restrictions on the data. Our approach, which is based on the free boundary analysis of Blanchet, Dolbeault and Monneau [5 ] and Blanchet [3 ] , shows that the first contact point ( τ , P τ ) (\tau,P_{\tau}) ( τ , P τ ) is almost surely regular on { 0 < τ < T } \{0<\tau<T\} { 0 < τ < T } ; the meaning of the term “regular” is explained later in the paper. This yields local C 0 , 1 / 2 C^{0,1/2} C 0 , 1/2 regularity of the contacted boundary, a key ingredient in establishing convergence of the optimal stopping times.
In our model, the initial belief being a random variable with distribution η \eta η models the heterogeneity of prior knowledge in the large scale system. This differs from the setting of [8 ] , where it is assumed that there is only common prior belief p 0 p_{0} p 0 .
Our objective is more general compared to [8 ] . We allow for time and state dependent running cost and a piecewise affine cost G G G without any additional monotonicity assumption. Moreover, we establish a quantitative finite population approximation, which is left open in [8 ] . Since the free boundary stability is the essential continuity input in mean field games of optimal stopping, our analysis provides a potential framework for studying a broader class of such problems.
Organization of the paper. In Section 2 we introduce the main assumptions, the mean field information acquisition model, the notion of a mean field solution and the main result about the existence of a mean field equilibrium as well as a number of illustrating examples. Section 3 is devoted to a detailed analysis of the underlying variational inequality and its stopping (free) boundaries. In Section 4 we establish the existence of a mean field equilibrium. Finally, in Section 5 we formulate and analyze a finite population game and show that the mean field equilibria obtained earlier are approximate Nash equilibria.
Preliminaries, assumptions and examples
We begin with some preliminaries about the weak topology of M \mathcal{M} M . Then we introduce the precise assumptions and conclude with a number of examples that fit our general setting.
The weak topology of M \mathcal{M} M
To deal with the convergence in M ( [ 0 , T ] ) \mathcal{M}([0,T]) M ([ 0 , T ]) , we adjoin a “cemetery point” † \dagger † to [ 0 , T ] [0,T] [ 0 , T ] , set I ‾ = [ 0 , T ] ∪ { † } \overline{I}=[0,T]\cup\{\dagger\} I = [ 0 , T ] ∪ { † } and introduce the distance d † d_{\dagger} d † on I ‾ \overline{I} I given by
d † ( s , r ) = ∣ s − r ∣ , d † ( s , † ) = d † ( † , s ) = T + 1 and d † ( † , † ) = 0. d_{\dagger}(s,r)=|s-r|,\qquad d_{\dagger}(s,\dagger)=d_{\dagger}(\dagger,s)=T+1\quad\text{and}\quad d_{\dagger}(\dagger,\dagger)=0. d † ( s , r ) = ∣ s − r ∣ , d † ( s , † ) = d † ( † , s ) = T + 1 and d † ( † , † ) = 0.
Let δ † \delta_{\dagger} δ † be the δ \delta δ −function at † \dagger † . We then extend a nonnegative subprobability measure λ \lambda λ on [ 0 , T ] [0,T] [ 0 , T ] to a probability measure λ ‾ \overline{\lambda} λ on I ‾ \overline{I} I by
λ ‾ : = λ + ( 1 − λ ( [ 0 , T ] ) ) δ † . \overline{\lambda}:=\lambda+\left(1-\lambda([0,T])\right)\delta_{\dagger}. λ := λ + ( 1 − λ ([ 0 , T ]) ) δ † .
In turn, the Wasserstein metric W 1 e x t W_1^{\mathrm{ext}} W 1 ext on M + ( [ 0 , T ] ) \mathcal{M}_+([0,T]) M + ([ 0 , T ]) is defined by
W 1 e x t ( λ , λ ′ ) : = W 1 ( λ ‾ , λ ′ ‾ ) . W_1^{\mathrm{ext}}(\lambda,\lambda'):=W_1(\overline{\lambda},\overline{\lambda'}). W 1 ext ( λ , λ ′ ) := W 1 ( λ , λ ′ ) .
We equip M \mathcal{M} M with the product weak topology metrized by
d M ( μ , ν ) : = W 1 e x t ( μ 0 , ν 0 ) + W 1 e x t ( μ 1 , ν 1 ) , d_{\mathcal{M}}(\mu,\nu):=W_1^{\mathrm{ext}}(\mu^0,\nu^0)+W_1^{\mathrm{ext}}(\mu^1,\nu^1), d M ( μ , ν ) := W 1 ext ( μ 0 , ν 0 ) + W 1 ext ( μ 1 , ν 1 ) ,
and note that it is compact and convex.
For each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , we consider the cumulative in time histories
μ t i : = μ i ( [ 0 , t ] ) , \mu_t^i:=\mu^i([0,t]), μ t i := μ i ([ 0 , t ]) ,
and, for convenience, we write
μ ≤ t i : = { μ s i : 0 ≤ s ≤ t } and μ ≤ t : = ( μ ≤ t 0 , μ ≤ t 1 ) . \mu_{\leq t}^i:=\{\mu_s^i:0\leq s\leq t\}\quad\text{and}\quad\mu_{\leq t}:=(\mu_{\leq t}^0,\mu_{\leq t}^1). μ ≤ t i := { μ s i : 0 ≤ s ≤ t } and μ ≤ t := ( μ ≤ t 0 , μ ≤ t 1 ) .
In the sequel, we will refer to μ ≤ t \mu_{\leq t} μ ≤ t as collective learning input or history.
Assumptions
To model the gradual, rather than instantaneous, effect of the past stopping-learning decisions we need to regularize μ ≤ t 0 \mu_{\leq t}^0 μ ≤ t 0 and μ ≤ t 1 \mu_{\leq t}^1 μ ≤ t 1 . For this, we select a nonnegative ϕ ∈ C c ∞ ( ( 0 , ∞ ) ) \phi\in C_c^\infty((0,\infty)) ϕ ∈ C c ∞ (( 0 , ∞ )) with ∫ R ϕ ( r ) d r = 1 \int_{\mathbb{R}}\phi(r)\,\mathrm{d}r=1 ∫ R ϕ ( r ) d r = 1 and introduce H ( r ) : = ∫ − ∞ r ϕ ( u ) d u H(r):=\int_{-\infty}^{r}\phi(u)\,\mathrm{d}u H ( r ) := ∫ − ∞ r ϕ ( u ) d u . Then we define, for i = 0 , 1 i=0,1 i = 0 , 1 ,
F ( t ; μ ≤ t ) : = ( F 0 ( t ; μ ≤ t 0 ) , F 1 ( t ; μ ≤ t 1 ) ) where F i ( t ; μ ≤ t i ) : = ∫ [ 0 , T ] H ( t − s ) μ i ( d s ) . (16) F(t;\mu_{\leq t}):=\left(F_0(t;\mu_{\leq t}^0),F_1(t;\mu_{\leq t}^1)\right)\quad\text{where}\quad F_i(t;\mu_{\leq t}^i):=\int_{[0,T]}H(t-s)\,\mu^i(\mathrm{d}s). \tag*{(16)} F ( t ; μ ≤ t ) := ( F 0 ( t ; μ ≤ t 0 ) , F 1 ( t ; μ ≤ t 1 ) ) where F i ( t ; μ ≤ t i ) := ∫ [ 0 , T ] H ( t − s ) μ i ( d s ) . (16)
Since F i ( t ; μ ≤ t i ) = ∫ 0 t μ s i ϕ ( t − s ) d s F_i(t;\mu_{\leq t}^i)=\int_0^t\mu_s^i\phi(t-s)\,\mathrm{d}s F i ( t ; μ ≤ t i ) = ∫ 0 t μ s i ϕ ( t − s ) d s , we see that F ( t ; μ ≤ t ) F(t;\mu_{\leq t}) F ( t ; μ ≤ t ) depends only on stopping decisions made before t t t , each of its components is nondecreasing in t t t , and F F F is [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 -valued and uniformly Lipschitz in t t t .
Fix μ ∈ M \mu\in\mathcal{M} μ ∈ M . For the signal drift b ( t , Θ ; F ( t ; μ ≤ t ) ) b(t,\Theta;F(t;\mu_{\leq t})) b ( t , Θ ; F ( t ; μ ≤ t )) in (3) , with Δ b ( t ; z ) : = b ( t , 1 ; z ) − b ( t , 0 ; z ) \Delta b(t;z):=b(t,1;z)-b(t,0;z) Δ b ( t ; z ) := b ( t , 1 ; z ) − b ( t , 0 ; z ) for z ∈ [ 0 , 1 ] 2 z\in[0,1]^2 z ∈ [ 0 , 1 ] 2 as in (5) , we assume that
b : [ 0 , T ] × { 0 , 1 } × [ 0 , 1 ] 2 → R is bounded and continuous , (17) b:[0,T]\times\{0,1\}\times[0,1]^2\to\mathbb{R}\text{ is bounded and continuous}, \tag*{(17)} b : [ 0 , T ] × { 0 , 1 } × [ 0 , 1 ] 2 → R is bounded and continuous , (17)
and
[ 0 , T ] × M ∋ ( t , μ ) ↦ Δ μ b ( t ) is Lipschitz continuous, and there exist constants 0 < c b ≤ c ‾ b < ∞ such that c b ≤ Δ μ b ( t ) ≤ c ‾ b uniformly in [ 0 , T ] × M . (18) [0,T]\times\mathcal{M}\ni(t,\mu)\mapsto\Delta^\mu b(t)\text{ is Lipschitz continuous, and there exist constants }0<c_b\leq\overline{c}_b<\infty\text{ such that }c_b\leq\Delta^\mu b(t)\leq\overline{c}_b\text{ uniformly in }[0,T]\times\mathcal{M}. \tag*{(18)} [ 0 , T ] × M ∋ ( t , μ ) ↦ Δ μ b ( t ) is Lipschitz continuous, and there exist constants 0 < c b ≤ c b < ∞ such that c b ≤ Δ μ b ( t ) ≤ c b uniformly in [ 0 , T ] × M . (18)
The boundedness and Lipschitz conditions on b b b and Δ b \Delta b Δ b ensure that the sde in (4) has a unique strong solution. Since 0 0 0 and 1 1 1 are absorbing states, P t , p ; μ P^{t,p;\mu} P t , p ; μ remains in [ 0 , 1 ] [0,1] [ 0 , 1 ] and is therefore a bounded F μ \mathbb{F}^{\mu} F μ -martingale.
It also follows from (18) that a μ a^\mu a μ defined in (11) is strictly bounded away from zero in compact subsets of ( 0 , 1 ) (0,1) ( 0 , 1 ) , since
a μ ( t , p ) = 1 2 ( Δ b μ ( t ) ) 2 p 2 ( 1 − p ) 2 ≥ 1 2 c b 2 p 2 ( 1 − p ) 2 . (19) a^\mu(t,p)=\frac{1}{2}\left(\Delta b^\mu(t)\right)^2p^2(1-p)^2\ge\frac{1}{2}c_b^2p^2(1-p)^2. \tag*{(19)} a μ ( t , p ) = 2 1 ( Δ b μ ( t ) ) 2 p 2 ( 1 − p ) 2 ≥ 2 1 c b 2 p 2 ( 1 − p ) 2 . (19)
For the running cost function introduced in (7) , we assume that f : [ 0 , T ] × [ 0 , 1 ] × [ 0 , 1 ] 2 → ( 0 , ∞ ) f:[0,T]\times[0,1]\times[0,1]^2\to(0,\infty) f : [ 0 , T ] × [ 0 , 1 ] × [ 0 , 1 ] 2 → ( 0 , ∞ ) satisfies
( t , p , z ) ↦ f ( t , p ; z ) is Lipschitz continuous, for every ( t , z ) the map p ↦ f ( t , p ; z ) is concave, and there exist constants 0 < c ‾ f ≤ C ‾ f < ∞ such that, for all ( t , p , z ) , c ‾ f ≤ f ( t , p ; z ) ≤ C ‾ f . (20) \begin{aligned}
(t,p,z)&\mapsto f(t,p;z)\text{ is Lipschitz continuous,}\\
&\text{for every }(t,z)\text{ the map }p\mapsto f(t,p;z)\text{ is concave, and}\\
&\text{there exist constants }0<\underline{c}_f\le\overline{C}_f<\infty\text{ such that, for all }(t,p,z),\\
\underline{c}_f&\le f(t,p;z)\le\overline{C}_f. \tag*{(20)}
\end{aligned} ( t , p , z ) c f ↦ f ( t , p ; z ) is Lipschitz continuous, for every ( t , z ) the map p ↦ f ( t , p ; z ) is concave, and there exist constants 0 < c f ≤ C f < ∞ such that, for all ( t , p , z ) , ≤ f ( t , p ; z ) ≤ C f . (20)
An admissible stopping-decision pair at time t t t consists of a [ t , T ] [t,T] [ t , T ] -valued F μ \mathcal{F}^\mu F μ -stopping time τ \tau τ and a F τ μ \mathcal{F}^\mu_\tau F τ μ -measurable decision d ∈ { 0 , 1 } d\in\{0,1\} d ∈ { 0 , 1 } . Here, d = 0 d=0 d = 0 and d = 1 d=1 d = 1 represent respectively the choice of the null hypothesis Θ = 0 \Theta=0 Θ = 0 , and of the alternative hypothesis Θ = 1 \Theta=1 Θ = 1 .
If learning terminates at time τ \tau τ with posterior belief P τ μ P^\mu_\tau P τ μ and a decision d d d is made, the resulting terminal loss G G G is the classical cost in sequential testing problems given by (8) .
Note that
G is a concave Lipschitz continuous function with constant a ˉ = max ( a 0 , a 1 ) . (21) G\text{ is a concave Lipschitz continuous function with constant }\bar{a}=\max(a_0,a_1). \tag*{(21)} G is a concave Lipschitz continuous function with constant a ˉ = max ( a 0 , a 1 ) . (21)
For some of the arguments that follow, it is more convenient to rewrite G G G as
G ( p ) = G ( m ) + a 0 ( p − m ) if p ≤ m and G ( p ) = G ( m ) − a 1 ( p − m ) if p ≥ m , G(p)=G(m)+a_0(p-m)\text{ if }p\le m\text{ and }G(p)=G(m)-a_1(p-m)\text{ if }p\ge m, G ( p ) = G ( m ) + a 0 ( p − m ) if p ≤ m and G ( p ) = G ( m ) − a 1 ( p − m ) if p ≥ m ,
with m m m given by (9) .
In order to make statements less cumbersome, we summarize the assumptions above in
assume that (1), (2), (17), (18), (20) and (9) hold. (22) \text{assume that (1), (2), (17), (18), (20) and (9) hold.} \tag*{(22)} assume that (1), (2), (17), (18), (20) and (9) hold. (22)
Examples within our framework
We provide three representative examples.
(i) Market liquidation. We consider a model for liquidity-driven information degradation in a financial market, where binary asset fundamentals are being learned from transaction prices. As trading stops and liquidation takes place, market depth deteriorates. For some constants β 0 > β 1 > 0 \beta_0>\beta_1>0 β 0 > β 1 > 0 , we model the informational signal to evolve according to
d Y t μ = Θ ( β 0 − β 1 ( F 0 ( t ; μ ≤ t μ ) + F 1 ( t ; μ ≤ t μ ) ) ) d t + d W t . \mathrm{d}Y^\mu_t=\Theta\left(\beta_0-\beta_1\left(F_0(t;\mu^\mu_{\le t})+F_1(t;\mu^\mu_{\le t})\right)\right)\mathrm{d}t+\mathrm{d}W_t. d Y t μ = Θ ( β 0 − β 1 ( F 0 ( t ; μ ≤ t μ ) + F 1 ( t ; μ ≤ t μ ) ) ) d t + d W t .
As liquidation accumulates, the drift gap between the two hypotheses narrows, slowing information acquisition by the remaining traders. A constant running cost f ( t , p ; μ ≤ t ) = c > 0 f(t,p;\mu_{\le t})=c>0 f ( t , p ; μ ≤ t ) = c > 0 represents the cost of continued monitoring, so aggregate liquidation feeds back into the value of waiting for more information.
(ii) Selective disclosure. We consider a review-driven learning with selective disclosure. Consumers decide whether a product is of high or low quality while reviews from earlier buyers affect both the signal and the learning running cost. For example, we may assume that, for some constants α , λ , κ > 0 \alpha, \lambda, \kappa> 0 α , λ , κ > 0 , the signal Y μ Y^\mu Y μ evolves according to
d Y t μ = α ( Θ + λ F 0 ( t ; μ ≤ t 0 ) − κ F 1 ( t ; μ ≤ t 1 ) ) d t + d W t . \mathrm{d}Y_t^\mu= \alpha\left(\Theta+ \lambda F_0(t;\mu^0_{\le t}) - \kappa F_1(t;\mu^1_{\le t})\right)\mathrm{d}t + \mathrm{d}W_t. d Y t μ = α ( Θ + λ F 0 ( t ; μ ≤ t 0 ) − κ F 1 ( t ; μ ≤ t 1 ) ) d t + d W t .
The aggregate reviews shift the observed signal level, while Δ b = α \Delta b = \alpha Δ b = α . A running cost such as
f ( t , p ; μ ≤ t ) = c 0 + c 1 ( F 0 ( t ; μ ≤ t 0 ) + F 1 ( t ; μ ≤ t 1 ) ) , f(t,p;\mu_{\le t}) = c_0 + c_1\left(F_0(t;\mu^0_{\le t}) + F_1(t;\mu^1_{\le t})\right), f ( t , p ; μ ≤ t ) = c 0 + c 1 ( F 0 ( t ; μ ≤ t 0 ) + F 1 ( t ; μ ≤ t 1 ) ) ,
where c 0 > 0 c_0 > 0 c 0 > 0 and c 1 ≥ 0 c_1 \ge0 c 1 ≥ 0 , captures declining attention or availability.
(iii) Information crowding. In experimentation platforms with A/B testing, many product teams experiment in parallel. However, widespread adoption reduces experimental diversity. We model, for some constants β , γ > 0 \beta, \gamma> 0 β , γ > 0 , the signal to evolve according to
d Y t μ = Θ β exp ( − γ ( F 0 ( t ; μ ≤ t 0 ) + F 1 ( t ; μ ≤ t 1 ) ) ) d t + d W t . \mathrm{d}Y_t^\mu= \Theta\beta\exp\left(-\gamma\left(F_0(t;\mu^0_{\le t}) + F_1(t;\mu^1_{\le t})\right)\right)\mathrm{d}t + \mathrm{d}W_t. d Y t μ = Θ β exp ( − γ ( F 0 ( t ; μ ≤ t 0 ) + F 1 ( t ; μ ≤ t 1 ) ) ) d t + d W t .
We then see that the drift decreases as more teams conclude their experiments. A cost
f ( t , p ; μ ≤ t ) = c 0 + c 1 t , f(t,p;\mu_{\le t}) = c_0 + c_1t, f ( t , p ; μ ≤ t ) = c 0 + c 1 t ,
with c 0 > 0 c_0 > 0 c 0 > 0 and c 1 ≥ 0 c_1 \ge0 c 1 ≥ 0 , represents the opportunity cost of delaying deployment. Among others, such feedback may generate early stopping and late experimentation equilibria.
The variational inequality and the free boundaries
This section is devoted to the analysis of the variational inequality, the properties of its solutions and their connection with the value function, as well as some important properties of the free boundaries.
Stability with respect to the collective input
In view of the particular dependence of the drift of the signal and the running cost on the collective input, we have the following result.
Proposition 3.1. If μ n ⇀ μ \mu^n \rightharpoonup\mu μ n ⇀ μ in M \mathcal{M} M , then
∥ Δ b μ n − Δ b μ ∥ L ∞ ( 0 , T ) + ∥ f ˉ μ n − f ˉ μ ∥ L ∞ ( [ 0 , T ] × [ 0 , 1 ] ) ⟶ 0. (23) \left\|\Delta b^{\mu^n}-\Delta b^\mu\right\|_{L^\infty(0,T)}+\left\|\bar f^{\mu^n}-\bar f^\mu\right\|_{L^\infty([0,T]\times[0,1])}\longrightarrow0. \tag*{(23)} Δ b μ n − Δ b μ L ∞ ( 0 , T ) + f ˉ μ n − f ˉ μ L ∞ ([ 0 , T ] × [ 0 , 1 ]) ⟶ 0. (23) Proof. For every t t t , the function s ↦ H ( t − s ) s\mapsto H(t-s) s ↦ H ( t − s ) on [ 0 , T ] [0,T] [ 0 , T ] , extended by the value 0 0 0 at † \dagger † , is Lipschitz on ( I , d † ) (I,d_\dagger) ( I , d † ) with constant at most max { ∥ ϕ ∥ ∞ , ( T + 1 ) − 1 } \max\{\|\phi\|_\infty,(T+1)^{-1}\} max { ∥ ϕ ∥ ∞ , ( T + 1 ) − 1 } .
It then follows from the Kantorovich–Rubinstein duality that
sup t ∈ [ 0 , T ] ∣ F ( t ; ( ( μ n ) ≤ t 0 , ( μ n ) ≤ t 1 ) ) − F ( t ; ( μ ≤ t 0 , μ ≤ t 1 ) ) ∣ ≤ max { ∥ ϕ ∥ ∞ , ( T + 1 ) − 1 } { W 1 e x t ( μ n , 0 , μ 0 ) + W 1 e x t ( μ n , 1 , μ 1 ) } ⟶ 0. \sup_{t\in[0,T]}\left|F\left(t;((\mu^n)^0_{\le t},(\mu^n)^1_{\le t})\right)-F\left(t;(\mu^0_{\le t},\mu^1_{\le t})\right)\right|
\le\max\{\|\phi\|_\infty,(T+1)^{-1}\}\left\{\mathcal{W}^{\mathrm{ext}}_1(\mu^{n,0},\mu^0)+\mathcal{W}^{\mathrm{ext}}_1(\mu^{n,1},\mu^1)\right\}\longrightarrow0. t ∈ [ 0 , T ] sup F ( t ; (( μ n ) ≤ t 0 , ( μ n ) ≤ t 1 ) ) − F ( t ; ( μ ≤ t 0 , μ ≤ t 1 ) ) ≤ max { ∥ ϕ ∥ ∞ , ( T + 1 ) − 1 } { W 1 ext ( μ n , 0 , μ 0 ) + W 1 ext ( μ n , 1 , μ 1 ) } ⟶ 0. The Lipschitz continuity of Δ b \Delta b Δ b and f ˉ \bar f f ˉ in the aggregate variable proves (23).
Viscosity solutions
We begin with the definition of the viscosity solutions to the terminal value problem (10) .
Definition 3.2. (i) A bounded continuous function u : [ 0 , T ] × [ 0 , 1 ] ↦ R u:[0,T]\times[0,1]\mapsto\mathbb{R} u : [ 0 , T ] × [ 0 , 1 ] ↦ R is a viscosity subsolution to (10) , if u ( T , ⋅ ) ≤ G ( ⋅ ) u(T,\cdot)\leq G(\cdot) u ( T , ⋅ ) ≤ G ( ⋅ ) on [ 0 , 1 ] [0,1] [ 0 , 1 ] and, for any test function ϕ ∈ C 1 , 2 ( [ 0 , T ] × [ 0 , 1 ] ) \phi\in C^{1,2}([0,T]\times[0,1]) ϕ ∈ C 1 , 2 ([ 0 , T ] × [ 0 , 1 ]) and any ( t 0 , p 0 ) ∈ ( 0 , T ) × [ 0 , 1 ] (t_{0},p_{0})\in(0,T)\times[0,1] ( t 0 , p 0 ) ∈ ( 0 , T ) × [ 0 , 1 ] , such that u − ϕ u-\phi u − ϕ attains a local maximum at ( t 0 , p 0 ) (t_{0},p_{0}) ( t 0 , p 0 ) with u ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) u(t_{0},p_{0})=\phi(t_{0},p_{0}) u ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) ,
max { − ∂ t ϕ ( t 0 , p 0 ) − a μ ( t 0 , p 0 ; μ ≤ t ) ∂ p 2 ϕ ( t 0 , p 0 ) − f ( t 0 , p 0 ; μ ≤ t ) , u ( t 0 , p 0 ) − G ( p 0 ) } ≤ 0. \max\left\{-\partial_{t}\phi(t_{0},p_{0})-a^{\mu}(t_{0},p_{0};\mu_{\leq t})\partial_{p}^{2}\phi(t_{0},p_{0})-f(t_{0},p_{0};\mu_{\leq t}),u(t_{0},p_{0})-G(p_{0})\right\}\leq0. max { − ∂ t ϕ ( t 0 , p 0 ) − a μ ( t 0 , p 0 ; μ ≤ t ) ∂ p 2 ϕ ( t 0 , p 0 ) − f ( t 0 , p 0 ; μ ≤ t ) , u ( t 0 , p 0 ) − G ( p 0 ) } ≤ 0. (ii) A bounded continuous function w : [ 0 , T ] × [ 0 , 1 ] → R w:[0,T]\times[0,1]\to\mathbb{R} w : [ 0 , T ] × [ 0 , 1 ] → R is a viscosity supersolution to (10) , if w ( T , ⋅ ) ≥ G ( ⋅ ) w(T,\cdot)\geq G(\cdot) w ( T , ⋅ ) ≥ G ( ⋅ ) on [ 0 , 1 ] [0,1] [ 0 , 1 ] and for any test function ϕ ∈ C 1 , 2 ( [ 0 , T ] × [ 0 , 1 ] ) \phi\in C^{1,2}([0,T]\times[0,1]) ϕ ∈ C 1 , 2 ([ 0 , T ] × [ 0 , 1 ]) and any ( t 0 , p 0 ) ∈ ( 0 , T ) × ( 0 , 1 ) (t_{0},p_{0})\in(0,T)\times(0,1) ( t 0 , p 0 ) ∈ ( 0 , T ) × ( 0 , 1 ) such that w − ϕ w-\phi w − ϕ attains a local minimum with w ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) w(t_{0},p_{0})=\phi(t_{0},p_{0}) w ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) ,
max { − ∂ t ϕ ( t 0 , p 0 ) − a μ ( t 0 , p 0 ; μ ≤ t ) ∂ p 2 ϕ ( t 0 , p 0 ) − f ( t 0 , p 0 ; μ ≤ t ) , w ( t 0 , p 0 ) − G ( p 0 ) } ≥ 0. \max\left\{-\partial_{t}\phi(t_{0},p_{0})-a^{\mu}(t_{0},p_{0};\mu_{\leq t})\partial_{p}^{2}\phi(t_{0},p_{0})-f(t_{0},p_{0};\mu_{\leq t}),w(t_{0},p_{0})-G(p_{0})\right\}\geq0. max { − ∂ t ϕ ( t 0 , p 0 ) − a μ ( t 0 , p 0 ; μ ≤ t ) ∂ p 2 ϕ ( t 0 , p 0 ) − f ( t 0 , p 0 ; μ ≤ t ) , w ( t 0 , p 0 ) − G ( p 0 ) } ≥ 0. (iii) A bounded continuous function v v v is a viscosity solution to (10) , if it is both a subsolution and a supersolution.
To simplify the exposition, we omit from now on the term “viscosity” with the understanding that all (subsolutions and supersolutions) are in the viscosity sense.
We note that, in view of the explicit form of a μ a^{\mu} a μ at p = 0 p=0 p = 0 and p = 1 p=1 p = 1 , no boundary conditions are needed, meaning that subsolutions and supersolutions to (10) are automatically subsolutions and supersolutions at p = 0 p=0 p = 0 and p = 1 p=1 p = 1 . This is the result of the next lemma.
Lemma 3.3. Fix μ ∈ M \mu\in\mathcal{M} μ ∈ M and let u , w ∈ C ( [ 0 , T ] × [ 0 , 1 ] ) u,w\in C([0,T]\times[0,1]) u , w ∈ C ([ 0 , T ] × [ 0 , 1 ]) be, respectively, a subsolution and a supersolution to (10) in ( 0 , T ) × ( 0 , 1 ) (0,T)\times(0,1) ( 0 , T ) × ( 0 , 1 ) . Then u u u is a subsolution and w w w is a supersolution to (10) in [ 0 , T ) × [ 0 , 1 ] [0,T)\times[0,1] [ 0 , T ) × [ 0 , 1 ] .
Proof. Since both claims are proved similarly, we only present the details for the subsolution case.
We also note that we only need to check what happens if, for a test function ϕ ∈ C 1 , 2 ( [ 0 , T ] × [ 0 , 1 ] ) \phi\in C^{1,2}([0,T]\times[0,1]) ϕ ∈ C 1 , 2 ([ 0 , T ] × [ 0 , 1 ]) , u − ϕ u-\phi u − ϕ attains a maximum at ( t 0 , p 0 ) (t_{0},p_{0}) ( t 0 , p 0 ) for t 0 ∈ [ 0 , T ) t_{0}\in[0,T) t 0 ∈ [ 0 , T ) and p 0 ∈ { 0 , 1 } p_{0}\in\{0,1\} p 0 ∈ { 0 , 1 } and u ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) u(t_{0},p_{0})=\phi(t_{0},p_{0}) u ( t 0 , p 0 ) = ϕ ( t 0 , p 0 ) .
We first consider the former case and note that, without any loss of generality, we may assume that the maximum is strict. If not, we replace ϕ \phi ϕ by ϕ ( t , p ) + K ( ( t − t 0 ) 2 + p 2 ) 2 \phi(t,p)+K((t-t_{0})^{2}+p^{2})^{2} ϕ ( t , p ) + K (( t − t 0 ) 2 + p 2 ) 2 for a sufficiently large K > 0 K>0 K > 0 .
Then, the map
( t , p ) ⟼ u ( t , p ) − ϕ ( t , p ) − ε ( 1 p + 1 t ) (t,p)\longmapsto u(t,p)-\phi(t,p)-\varepsilon\left(\frac{1}{p}+\frac{1}{t}\right) ( t , p ) ⟼ u ( t , p ) − ϕ ( t , p ) − ε ( p 1 + t 1 )
attains a maximum at some ( t ε , p ε ) ∈ ( 0 , T ) × ( 0 , 1 ) (t_{\varepsilon},p_{\varepsilon})\in(0,T)\times(0,1) ( t ε , p ε ) ∈ ( 0 , T ) × ( 0 , 1 ) .
It then follows from the continuity of u u u and the strictness of the maximum above that, as ε → 0 \varepsilon\to0 ε → 0 ,
( t ε , p ε ) ⟶ ( t 0 , 0 ) and ε ( 1 p ε + 1 t ε ) ⟶ 0. (24) (t_{\varepsilon},p_{\varepsilon})\longrightarrow(t_{0},0)\quad\text{and}\quad\varepsilon\left(\frac{1}{p_{\varepsilon}}+\frac{1}{t_{\varepsilon}}\right)\longrightarrow0. \tag*{(24)} ( t ε , p ε ) ⟶ ( t 0 , 0 ) and ε ( p ε 1 + t ε 1 ) ⟶ 0. (24)
Since u u u is a subsolution in ( 0 , T ) × ( 0 , 1 ) (0,T)\times(0,1) ( 0 , T ) × ( 0 , 1 ) , it follows that
max { − ∂ t ϕ ( t ε , p ε ) + ε t ε 2 − a μ ( t ε , p ε ; μ ≤ t ε ) ( ∂ p 2 ϕ ( t ε , p ε ) + 2 ε p ε 3 ) − f ( t ε , p ε ; μ ≤ t ε ) , u ( t ε , p ε ) − G ( p ε ) } ≤ 0. \max\left\{-\partial_{t}\phi(t_{\varepsilon},p_{\varepsilon})+\frac{\varepsilon}{t_{\varepsilon}^{2}}-a^{\mu}(t_{\varepsilon},p_{\varepsilon};\mu_{\leq t_{\varepsilon}})\left(\partial_{p}^{2}\phi(t_{\varepsilon},p_{\varepsilon})+\frac{2\varepsilon}{p_{\varepsilon}^{3}}\right)-f(t_{\varepsilon},p_{\varepsilon};\mu_{\leq t_{\varepsilon}}),u(t_{\varepsilon},p_{\varepsilon})-G(p_{\varepsilon})\right\}\leq0. max { − ∂ t ϕ ( t ε , p ε ) + t ε 2 ε − a μ ( t ε , p ε ; μ ≤ t ε ) ( ∂ p 2 ϕ ( t ε , p ε ) + p ε 3 2 ε ) − f ( t ε , p ε ; μ ≤ t ε ) , u ( t ε , p ε ) − G ( p ε ) } ≤ 0.
Using (11) , we find that
a μ ( t ε , p ε ; μ ≤ t ε ) 2 ε p ε 3 = ε p ε ( Δ b ( t ε ; μ ≤ t ε ) ) 2 ( 1 − p ε ) 2 , a^\mu(t_\varepsilon,p_\varepsilon;\mu_{\le t_\varepsilon})\frac{2\varepsilon}{p_\varepsilon^3}=\frac{\varepsilon}{p_\varepsilon}\left(\Delta b(t_\varepsilon;\mu_{\le t_\varepsilon})\right)^2(1-p_\varepsilon)^2, a μ ( t ε , p ε ; μ ≤ t ε ) p ε 3 2 ε = p ε ε ( Δ b ( t ε ; μ ≤ t ε ) ) 2 ( 1 − p ε ) 2 ,
and, hence,
lim ε → 0 a μ ( t ε , p ε ; μ ≤ t ε ) 2 ε p ε 3 = 0. (25) \lim_{\varepsilon\to0}a^\mu(t_\varepsilon,p_\varepsilon;\mu_{\le t_\varepsilon})\frac{2\varepsilon}{p_\varepsilon^3}=0. \tag*{(25)} ε → 0 lim a μ ( t ε , p ε ; μ ≤ t ε ) p ε 3 2 ε = 0. (25)
Moreover, by the continuity of the coefficients, we have
lim ε → 0 a μ ( t ε , p ε ; μ ≤ t ε ) ∂ p 2 ϕ ( t ε , p ε ) = a μ ( t 0 , 0 ; μ ≤ t 0 ) ∂ p 2 ϕ ( t 0 , 0 ) , lim ε → 0 f ( t ε , p ε ; μ ≤ t ε ) = f ( t 0 , 0 ; μ ≤ t 0 ) . (26) \begin{aligned}
\lim_{\varepsilon\to0}a^\mu(t_\varepsilon,p_\varepsilon;\mu_{\le t_\varepsilon})\partial_p^2\phi(t_\varepsilon,p_\varepsilon)&=a^\mu(t_0,0;\mu_{\le t_0})\partial_p^2\phi(t_0,0),\\
\lim_{\varepsilon\to0}f(t_\varepsilon,p_\varepsilon;\mu_{\le t_\varepsilon})&=f(t_0,0;\mu_{\le t_0}). \tag*{(26)}
\end{aligned} ε → 0 lim a μ ( t ε , p ε ; μ ≤ t ε ) ∂ p 2 ϕ ( t ε , p ε ) ε → 0 lim f ( t ε , p ε ; μ ≤ t ε ) = a μ ( t 0 , 0 ; μ ≤ t 0 ) ∂ p 2 ϕ ( t 0 , 0 ) , = f ( t 0 , 0 ; μ ≤ t 0 ) . (26)
Using (24) , (25) and (26) , and omitting the nonnegative term ε / t ε 2 \varepsilon/t_\varepsilon^2 ε / t ε 2 , we obtain the sub-solution inequality at ( t 0 , 0 ) (t_0,0) ( t 0 , 0 ) .
If the maximum occurs at p 0 = 1 p_0=1 p 0 = 1 , we repeat the same argument the only difference being that we use 1 1 − p \frac{1}{1-p} 1 − p 1 in place of 1 p \frac{1}{p} p 1 .
The following result is well known and we refer to Crandall, Ishii and Lions [13 ] for its proof.
Theorem 3.4. Let u u u and w w w be respectively a bounded continuous subsolution and a bounded continuous supersolution to (10) , and assume that u ( T , ⋅ ) ≤ w ( T , ⋅ ) u(T,\cdot)\le w(T,\cdot) u ( T , ⋅ ) ≤ w ( T , ⋅ ) on [ 0 , 1 ] [0,1] [ 0 , 1 ] . Then, u ≤ w u\le w u ≤ w on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] . In particular, (10) has a unique continuous viscosity solution.
Next, we argue that, for each μ ∈ M \mu\in\mathcal{M} μ ∈ M , the value function V μ V^\mu V μ is the unique continuous viscosity solution to (10) . A key fact towards establishing this claim is that the value function satisfies the dynamic programming principle, which states (see Fleming and Soner [20 Chapters V and VIII] ) that, for every θ ∈ T t , T \theta\in\mathcal{T}_{t,T} θ ∈ T t , T ,
V μ ( t , p ) = inf τ ∈ T t , T E [ ∫ t τ ∧ θ f μ ( s , P s t , p ; μ ) d s + V μ ( τ ∧ θ , P τ ∧ θ t , p ; μ ) ] . (27) V^\mu(t,p)=\inf_{\tau\in\mathcal{T}_{t,T}}\mathbb{E}\left[\int_t^{\tau\wedge\theta}f^\mu(s,P_s^{t,p;\mu})\,\mathrm{d}s+V^\mu(\tau\wedge\theta,P_{\tau\wedge\theta}^{t,p;\mu})\right]. \tag*{(27)} V μ ( t , p ) = τ ∈ T t , T inf E [ ∫ t τ ∧ θ f μ ( s , P s t , p ; μ ) d s + V μ ( τ ∧ θ , P τ ∧ θ t , p ; μ ) ] . (27)
We continue with the space-time continuity property of V μ V^\mu V μ .
Proposition 3.5. Assume (22) and fix μ ∈ M \mu\in\mathcal{M} μ ∈ M . For each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , V μ ( t , 0 ) = G ( 0 ) V^\mu(t,0)=G(0) V μ ( t , 0 ) = G ( 0 ) and V μ ( t , 1 ) = G ( 1 ) V^\mu(t,1)=G(1) V μ ( t , 1 ) = G ( 1 ) and, for each p ∈ [ 0 , 1 ] p\in[0,1] p ∈ [ 0 , 1 ] , V μ ( T , p ) = G ( p ) V^\mu(T,p)=G(p) V μ ( T , p ) = G ( p ) . Moreover, V μ V^\mu V μ is Lipschitz continuous in space and 1 / 2 1/2 1/2 –Hölder continuous in time on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] .
Proof. If p = 0 p=0 p = 0 , then P s t , 0 ; μ = 0 P_s^{t,0;\mu}=0 P s t , 0 ; μ = 0 for all 0 ≤ t ≤ s ≤ T 0\le t\le s\le T 0 ≤ t ≤ s ≤ T . Consequently, for every τ ∈ T t , T \tau\in\mathcal{T}_{t,T} τ ∈ T t , T ,
E [ ∫ t τ ∧ T f μ ( s , 0 ) d s + G ( 0 ) ] ≥ G ( 0 ) , \mathbb{E}\left[\int_t^{\tau\wedge T}f^\mu(s,0)\,\mathrm{d}s+G(0)\right]\ge G(0), E [ ∫ t τ ∧ T f μ ( s , 0 ) d s + G ( 0 ) ] ≥ G ( 0 ) ,
and equality is attained by τ = t \tau=t τ = t . The proof at p = 1 p=1 p = 1 is similar. When t = T t=T t = T , the only admissible stopping time is T T T , and, hence, V μ ( T , p ) = G ( p ) V^\mu(T,p)=G(p) V μ ( T , p ) = G ( p ) for all p ∈ [ 0 , 1 ] p\in[0,1] p ∈ [ 0 , 1 ] .
To prove the Lipschitz continuity in space, take t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] and p , q ∈ [ 0 , 1 ] p,q\in[0,1] p , q ∈ [ 0 , 1 ] , and recall that P t , p ; μ P^{t,p;\mu} P t , p ; μ and P t , q ; μ P^{t,q;\mu} P t , q ; μ solve (4) with the same Brownian motion. Since, in view of (11) and (19) , the diffusion coefficient is uniformly Lipschitz in the state variable away from { 0 , 1 } \{0,1\} { 0 , 1 } , the Burkholder–Davis–Grundy and Grönwall inequalities give
E [ sup t ≤ r ≤ T ∣ P r t , p ; μ − P r t , q ; μ ∣ 2 ] ≤ C ∣ p − q ∣ 2 . \mathbb{E}\left[\sup_{t\le r\le T}\left|P_{r}^{t,p;\mu}-P_{r}^{t,q;\mu}\right|^{2}\right]\le C|p-q|^{2}. E [ t ≤ r ≤ T sup P r t , p ; μ − P r t , q ; μ 2 ] ≤ C ∣ p − q ∣ 2 .
Hence, for every τ ∈ T t , T μ \tau\in\mathcal{T}_{t,T}^{\mu} τ ∈ T t , T μ ,
∣ E [ ∫ t τ f μ ( r , P r t , p ; μ ) d r + G ( P τ t , p ; μ ) ] − E [ ∫ t τ f μ ( r , P r t , q ; μ ) d r + G ( P τ t , q ; μ ) ] ∣ ≤ L f E ∫ t τ ∣ P r t , p ; μ − P r t , q ; μ ∣ d r + a ˉ E ∣ P τ t , p ; μ − P τ t , q ; μ ∣ ≤ C ∣ p − q ∣ , \left|\mathbb{E}\left[\int_{t}^{\tau}f^{\mu}(r,P_{r}^{t,p;\mu})\,\mathrm{d}r+G(P_{\tau}^{t,p;\mu})\right]-\mathbb{E}\left[\int_{t}^{\tau}f^{\mu}(r,P_{r}^{t,q;\mu})\,\mathrm{d}r+G(P_{\tau}^{t,q;\mu})\right]\right|
\le L_f\mathbb{E}\int_{t}^{\tau}\left|P_{r}^{t,p;\mu}-P_{r}^{t,q;\mu}\right|\,\mathrm{d}r+\bar a\mathbb{E}\left|P_{\tau}^{t,p;\mu}-P_{\tau}^{t,q;\mu}\right|\le C|p-q|, E [ ∫ t τ f μ ( r , P r t , p ; μ ) d r + G ( P τ t , p ; μ ) ] − E [ ∫ t τ f μ ( r , P r t , q ; μ ) d r + G ( P τ t , q ; μ ) ] ≤ L f E ∫ t τ P r t , p ; μ − P r t , q ; μ d r + a ˉ E P τ t , p ; μ − P τ t , q ; μ ≤ C ∣ p − q ∣ ,
where a ˉ \bar a a ˉ is as in (21) .
Taking infima over τ ∈ T t , T μ \tau\in\mathcal{T}_{t,T}^{\mu} τ ∈ T t , T μ in the above estimates gives
∣ V μ ( t , p ) − V μ ( t , q ) ∣ ≤ C ∣ p − q ∣ . (28) \left|V^{\mu}(t,p)-V^{\mu}(t,q)\right|\le C|p-q|. \tag*{(28)} ∣ V μ ( t , p ) − V μ ( t , q ) ∣ ≤ C ∣ p − q ∣. (28)
For the continuity in time, we fix 0 ≤ t < s ≤ T 0\le t<s\le T 0 ≤ t < s ≤ T , use (27) with θ = s \theta=s θ = s and choose an ε \varepsilon ε –approximate minimizing stopping time τ ε \tau^{\varepsilon} τ ε for V μ ( t , p ) V^{\mu}(t,p) V μ ( t , p ) . Using that by, its definition, V μ ( s , p ) ≤ G ( p ) V^{\mu}(s,p)\le G(p) V μ ( s , p ) ≤ G ( p ) , the positivity of f μ f^{\mu} f μ , the Lipschitz continuity in space of G G G and V μ V^{\mu} V μ and the fact that, for some c > 0 c>0 c > 0 ,
E [ ∣ P ρ t , p ; μ − p ∣ ] ≤ c ∣ ρ − t ∣ , \mathbb{E}\left[\left|P_{\rho}^{t,p;\mu}-p\right|\right]\le c\sqrt{|\rho-t|}, E [ P ρ t , p ; μ − p ] ≤ c ∣ ρ − t ∣ ,
we find
V ( t , p ) − V ( s , p ) ≥ E [ ∫ t τ ε ∧ s f μ ( ρ , P ρ t , p ; μ ) d ρ + 1 { τ ε < s } G ( P τ ε t , p ; μ ) + 1 { τ ε ≥ s } V μ ( θ , P s t , p ; μ ) − V μ ( s , p ) ] − ε ≥ E [ 1 { τ ε < s } ( G ( P τ ε t , p ; μ ) − G ( p ) ) ] + E [ 1 { τ ε ≥ θ } ( V μ ( s , P s t , p ; μ ) − V μ ( s , p ) ) ] − ε ≥ − a ˉ E [ 1 { τ ε ≤ s } ∣ P τ ε t , p ; μ − p ∣ ] − C E [ ∣ P s t , p ; μ − p ∣ ] − ε ≥ − c ( a ˉ + C ) ∣ s − t ∣ − ε . \begin{aligned}
V(t,p)-V(s,p)&\ge\mathbb{E}\left[\int_{t}^{\tau^{\varepsilon}\wedge s}f^{\mu}(\rho,P_{\rho}^{t,p;\mu})\,\mathrm{d}\rho+\mathbf{1}_{\{\tau^{\varepsilon}<s\}}G(P_{\tau^{\varepsilon}}^{t,p;\mu})\right.\\
&\qquad\left.+\mathbf{1}_{\{\tau^{\varepsilon}\ge s\}}V^{\mu}(\theta,P_{s}^{t,p;\mu})-V^{\mu}(s,p)\right]-\varepsilon\\
&\ge\mathbb{E}\left[\mathbf{1}_{\{\tau^{\varepsilon}<s\}}\left(G(P_{\tau^{\varepsilon}}^{t,p;\mu})-G(p)\right)\right]+\mathbb{E}\left[\mathbf{1}_{\{\tau^{\varepsilon}\ge\theta\}}\left(V^{\mu}(s,P_{s}^{t,p;\mu})-V^{\mu}(s,p)\right)\right]-\varepsilon\\
&\ge-\bar a\mathbb{E}\left[\mathbf{1}_{\{\tau^{\varepsilon}\le s\}}\left|P_{\tau^{\varepsilon}}^{t,p;\mu}-p\right|\right]-C\mathbb{E}\left[\left|P_{s}^{t,p;\mu}-p\right|\right]-\varepsilon\\
&\ge-c(\bar a+C)\sqrt{|s-t|}-\varepsilon.
\end{aligned} V ( t , p ) − V ( s , p ) ≥ E [ ∫ t τ ε ∧ s f μ ( ρ , P ρ t , p ; μ ) d ρ + 1 { τ ε < s } G ( P τ ε t , p ; μ ) + 1 { τ ε ≥ s } V μ ( θ , P s t , p ; μ ) − V μ ( s , p ) ] − ε ≥ E [ 1 { τ ε < s } ( G ( P τ ε t , p ; μ ) − G ( p ) ) ] + E [ 1 { τ ε ≥ θ } ( V μ ( s , P s t , p ; μ ) − V μ ( s , p ) ) ] − ε ≥ − a ˉ E [ 1 { τ ε ≤ s } P τ ε t , p ; μ − p ] − C E [ P s t , p ; μ − p ] − ε ≥ − c ( a ˉ + C ) ∣ s − t ∣ − ε .
A similar argument, choosing now an ε \varepsilon ε –minimizing stopping time for V μ ( s , p ) V^{\mu}(s,p) V μ ( s , p ) leads, for some C ′ > 0 C'>0 C ′ > 0 , to
V ( s , p ) − V ( t , p ) ≥ − C ′ ( L f ∣ s − t ∣ + ( a ˉ + C ) ∣ s − t ∣ 1 / 2 ) − ε . V(s,p)-V(t,p)\ge-C'\left(L_f|s-t|+(\bar a+C)|s-t|^{1/2}\right)-\varepsilon. V ( s , p ) − V ( t , p ) ≥ − C ′ ( L f ∣ s − t ∣ + ( a ˉ + C ) ∣ s − t ∣ 1/2 ) − ε .
It then follows, after letting ε → 0 \varepsilon\to0 ε → 0 , that
∣ V ( s , p ) − V ( t , p ) ∣ ≲ ∣ s − t ∣ + ∣ s − t ∣ . (29) |V(s,p)-V(t,p)|\lesssim|s-t|+\sqrt{|s-t|}. \tag*{(29)} ∣ V ( s , p ) − V ( t , p ) ∣ ≲ ∣ s − t ∣ + ∣ s − t ∣ . (29)
We explore next the connection between V μ V^{\mu} V μ and the solution to (10) .
Proposition 3.6. Assume (22). For any μ ∈ M \mu\in\mathcal{M} μ ∈ M , V μ V^{\mu} V μ is the unique continuous viscosity solution to (10).
Proof. In view of Theorem 3.4 and Proposition 3.5, we only need to check that V μ V^\mu V μ is a viscosity solution.
It follows from (27) that, for every θ ∈ T t , T \theta\in\mathcal{T}_{t,T} θ ∈ T t , T ,
V μ ( t , p ) = inf τ ∈ T t , T E [ ∫ t τ ∧ θ f μ ( s , P s t , p ; μ ) d s + 1 { τ < θ } G ( P τ t , p ; μ ) + 1 { τ ≥ θ } V μ ( θ , P θ t , p ; μ ) ] . V^\mu(t,p)=\inf_{\tau\in\mathcal{T}_{t,T}}\mathbb{E}\left[\int_t^{\tau\wedge\theta}f^\mu(s,P_s^{t,p;\mu})\,\mathrm{d}s+\mathbf{1}_{\{\tau<\theta\}}G(P_\tau^{t,p;\mu})+\mathbf{1}_{\{\tau\ge\theta\}}V^\mu(\theta,P_\theta^{t,p;\mu})\right]. V μ ( t , p ) = τ ∈ T t , T inf E [ ∫ t τ ∧ θ f μ ( s , P s t , p ; μ ) d s + 1 { τ < θ } G ( P τ t , p ; μ ) + 1 { τ ≥ θ } V μ ( θ , P θ t , p ; μ ) ] .
To obtain the viscosity inequalities at ( t , p ) (t,p) ( t , p ) , we introduce, for h , r > 0 h,r>0 h , r > 0 small, the stopping time
θ h : = ( t + h ) ∧ inf { s ≥ t : ∣ P s t , p ; μ − p ∣ ≥ r } . \theta_h:=(t+h)\wedge\inf\{s\ge t:\lvert P_s^{t,p;\mu}-p\rvert\ge r\}. θ h := ( t + h ) ∧ inf { s ≥ t : ∣ P s t , p ; μ − p ∣ ≥ r } .
If a smooth function ϕ \phi ϕ touches V μ V^\mu V μ from above at ( t , p ) (t,p) ( t , p ) , using the dynamic programming principle and Itô’s formula gives
− ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≤ 0. -\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu\le0. − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≤ 0.
Since immediate stopping gives V μ − G ≤ 0 V^\mu-G\le0 V μ − G ≤ 0 , we obtain
max { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ , V μ − G } ( t , p ) ≤ 0. \max\{-\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu,V^\mu-G\}(t,p)\le0. max { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ , V μ − G } ( t , p ) ≤ 0.
If ϕ \phi ϕ touches V μ V^\mu V μ from below at ( t , p ) (t,p) ( t , p ) , then
max { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ , V μ − G } ( t , p ) ≥ { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≥ 0 if V μ ( t , p ) < G ( p ) , 0 if V μ ( t , p ) = G ( p ) . \max\{-\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu,V^\mu-G\}(t,p)\ge
\begin{cases}
-\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu\ge0 & \text{if } V^\mu(t,p)<G(p),\\
0 & \text{if } V^\mu(t,p)=G(p).
\end{cases} max { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ , V μ − G } ( t , p ) ≥ { − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≥ 0 0 if V μ ( t , p ) < G ( p ) , if V μ ( t , p ) = G ( p ) .
The next result is an immediate consequence of the stability of viscosity solutions and the continuity of the coefficient of (10) with respect to the weak convergence in M \mathcal{M} M .
Proposition 3.7. If μ n ⇀ μ \mu^n\rightharpoonup\mu μ n ⇀ μ in M \mathcal{M} M , then
∥ V μ n − V μ ∥ L ∞ ( [ 0 , T ] × [ 0 , 1 ] ) ⟶ 0. \lVert V^{\mu^n}-V^\mu\rVert_{L^\infty([0,T]\times[0,1])}\longrightarrow0. ∥ V μ n − V μ ∥ L ∞ ([ 0 , T ] × [ 0 , 1 ]) ⟶ 0.
Proof. We first observe that the estimates (28) and (29) are uniform in μ \mu μ . Hence, the sequence ( V μ n ) n (V^{\mu^n})_n ( V μ n ) n is uniformly bounded and equicontinuous on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] . Let ( V μ n k ) k (V^{\mu^{n_k}})_k ( V μ n k ) k be any subsequence. By the Arzelà–Ascoli theorem, it has a further subsequence converging uniformly to a continuous function V ‾ \overline{V} V .
Since, as claimed in Proposition 3.1,
∥ a μ n k − a μ ∥ ∞ and ∥ f μ n k − f μ ∥ ∞ ⟶ 0 , \lVert a^{\mu^{n_k}}-a^\mu\rVert_\infty\quad\text{and}\quad\lVert f^{\mu^{n_k}}-f^\mu\rVert_\infty\longrightarrow0, ∥ a μ n k − a μ ∥ ∞ and ∥ f μ n k − f μ ∥ ∞ ⟶ 0 ,
the stability of viscosity solutions implies that V ‾ \overline{V} V solves (10) with the same terminal data. Then, in view of the uniqueness in Proposition 3.6, we have V ‾ = V μ \overline{V}=V^\mu V = V μ . Thus, every subsequence has a further subsequence converging uniformly to V μ V^\mu V μ , and we easily conclude.
Concavity of the value function in the space variable
To establish the concavity of the value function in the belief variable, we consider a standard penalization of the variational inequality. The reason is that it provides an equation instead of a variational inequality which can be, in turn, differentiated to obtain estimates.
Let β ∈ C ∞ ( R ) \beta\in C^\infty(\mathbb{R}) β ∈ C ∞ ( R ) be
convex and nondecreasing, with
β ( r ) = 0 for r ≤ 0 , β ( r ) > 0 for r > 0 , lim r → ∞ β ( r ) = ∞ and β ( 1 ) ≥ C f , (30) \beta(r)=0 \quad\text{for } r\leq0,\qquad\beta(r)>0 \quad\text{for } r>0,\qquad\lim_{r\to\infty}\beta(r)=\infty\quad\text{and}\quad\beta(1)\geq C_f, \tag*{(30)} β ( r ) = 0 for r ≤ 0 , β ( r ) > 0 for r > 0 , r → ∞ lim β ( r ) = ∞ and β ( 1 ) ≥ C f , (30)
and, for ε > 0 \varepsilon>0 ε > 0 , set
β ε ( r ) : = β ( r ε ) . (31) \beta_\varepsilon(r):=\beta\left(\frac{r}{\varepsilon}\right). \tag*{(31)} β ε ( r ) := β ( ε r ) . (31)
For fixed μ ∈ M \mu\in\mathcal{M} μ ∈ M , we then consider the auxiliary problem
{ − ∂ t V μ , ε − a μ ∂ p 2 V μ , ε − f μ + β ε ( V μ , ε − G ) = 0 in [ 0 , T ) × [ 0 , 1 ] , V μ , ε ( T , ⋅ ) = G on [ 0 , 1 ] . (32) \begin{cases}
-\partial_t V^{\mu,\varepsilon}-a^\mu\partial_p^2 V^{\mu,\varepsilon}-f^\mu+\beta_\varepsilon(V^{\mu,\varepsilon}-G)=0 & \text{in } [0,T)\times[0,1],\\
V^{\mu,\varepsilon}(T,\cdot)=G & \text{on } [0,1]. \tag*{(32)}
\end{cases} { − ∂ t V μ , ε − a μ ∂ p 2 V μ , ε − f μ + β ε ( V μ , ε − G ) = 0 V μ , ε ( T , ⋅ ) = G in [ 0 , T ) × [ 0 , 1 ] , on [ 0 , 1 ] . (32)
At p = 0 , 1 p=0,1 p = 0 , 1 , (32) is understood in the viscosity sense relative to the closed state space without imposing any boundary conditions.
The following result is well known, so we only provide a sketch of its proof.
Proposition 3.8. If 0 < ε ′ < ε 0<\varepsilon'<\varepsilon 0 < ε ′ < ε , then
V μ ≤ V μ , ε ′ ≤ V μ , ε on [ 0 , T ] × [ 0 , 1 ] , V^\mu\leq V^{\mu,\varepsilon'}\leq V^{\mu,\varepsilon}\quad\text{on }[0,T]\times[0,1], V μ ≤ V μ , ε ′ ≤ V μ , ε on [ 0 , T ] × [ 0 , 1 ] , and
∥ V μ , ε − V μ ∥ L ∞ ( [ 0 , T ] × [ 0 , 1 ] ) ≤ ε . \lVert V^{\mu,\varepsilon}-V^\mu\rVert_{L^\infty([0,T]\times[0,1])}\leq\varepsilon. ∥ V μ , ε − V μ ∥ L ∞ ([ 0 , T ] × [ 0 , 1 ]) ≤ ε .
Proof. Since β ε \beta_\varepsilon β ε is nondecreasing, the standard viscosity existence theorem gives a bounded continuous solution of (32), while the comparison principle yields uniqueness; see [13 Theorem 4.1 and Remark 4.5] .
Since β ε ′ ( r ) ≥ β ε ( r ) \beta_{\varepsilon'}(r)\geq\beta_\varepsilon(r) β ε ′ ( r ) ≥ β ε ( r ) if 0 < ε ′ < ε 0<\varepsilon'<\varepsilon 0 < ε ′ < ε , it follows that V μ , ε V^{\mu,\varepsilon} V μ , ε is a viscosity supersolution of the ε ′ \varepsilon' ε ′ -equation, and, thus,
V μ , ε ′ ≤ V μ , ε . V^{\mu,\varepsilon'}\leq V^{\mu,\varepsilon}. V μ , ε ′ ≤ V μ , ε .
We next compare V μ , ε V^{\mu,\varepsilon} V μ , ε with V μ V^\mu V μ noting that, in view of the fact that V μ ≤ G V^\mu\leq G V μ ≤ G , (30) gives β ε ( V μ − G ) = 0 \beta_\varepsilon(V^\mu-G)=0 β ε ( V μ − G ) = 0 . Using that
− ∂ t V μ − a μ ∂ p 2 V μ − f μ ≤ 0 in [ 0 , T ) × [ 0 , 1 ] , -\partial_tV^\mu-a^\mu\partial_p^2V^\mu-f^\mu\leq0\quad\text{in }[0,T)\times[0,1], − ∂ t V μ − a μ ∂ p 2 V μ − f μ ≤ 0 in [ 0 , T ) × [ 0 , 1 ] ,
we deduce that V μ V^\mu V μ is a sub-solution to (32) and, thus, for each ε > 0 \varepsilon>0 ε > 0 ,
V μ ≤ V μ , ε on [ 0 , T ] × [ 0 , 1 ] . V^\mu\leq V^{\mu,\varepsilon}\quad\text{on }[0,T]\times[0,1]. V μ ≤ V μ , ε on [ 0 , T ] × [ 0 , 1 ] .
Next, we claim that U = V μ + ε U=V^\mu+\varepsilon U = V μ + ε is a supersolution to (32), a fact, which in combination with the previous bound, yields (33).
Assume that, for a smooth ϕ \phi ϕ , U − ϕ U-\phi U − ϕ attains a local minimum at some ( t 0 , p 0 ) ∈ [ 0 , T ) × ( 0 , 1 ) (t_0,p_0)\in[0,T)\times(0,1) ( t 0 , p 0 ) ∈ [ 0 , T ) × ( 0 , 1 ) . It follows that, at ( t 0 , p 0 ) (t_0,p_0) ( t 0 , p 0 ) ,
− ∂ t ϕ − a μ ∂ p 2 ϕ − f μ + β ε ( U − G ) ≥ 0 , (33) -\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu+\beta_\varepsilon(U-G)\geq0, \tag*{(33)} − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ + β ε ( U − G ) ≥ 0 , (33)
and, thus, the claim.
Indeed, if V μ ( t 0 , p 0 ) < G ( p 0 ) V^\mu(t_0,p_0)<G(p_0) V μ ( t 0 , p 0 ) < G ( p 0 ) , then, since V μ − ϕ V^\mu-\phi V μ − ϕ attains a minimum at ( t 0 , p 0 ) (t_0,p_0) ( t 0 , p 0 ) , the supersolution property of V μ V^\mu V μ gives that, at ( t 0 , p 0 ) (t_0,p_0) ( t 0 , p 0 ) ,
− ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≥ 0 , -\partial_t\phi-a^\mu\partial_p^2\phi-f^\mu\geq0, − ∂ t ϕ − a μ ∂ p 2 ϕ − f μ ≥ 0 ,
and, in view of the nonnegativity of β ε \beta_\varepsilon β ε , (33) follows.
If V μ ( t 0 , p 0 ) = G ( p 0 ) V^\mu(t_0,p_0)=G(p_0) V μ ( t 0 , p 0 ) = G ( p 0 ) , then, we have that ( t 0 , p 0 ) (t_0,p_0) ( t 0 , p 0 ) is a minimum of G − ϕ G-\phi G − ϕ . It follows that p 0 ≠ m p_0\ne m p 0 = m and, in addition,
∂ t ϕ ( t 0 , p 0 ) − a μ ∂ p 2 ϕ ( t 0 , p 0 ) ≥ 0. \partial_t\phi(t_0,p_0)-a^\mu\partial_p^2\phi(t_0,p_0)\geq0. ∂ t ϕ ( t 0 , p 0 ) − a μ ∂ p 2 ϕ ( t 0 , p 0 ) ≥ 0.
Moreover, in view of (30) , at ( t 0 , p 0 ) (t_0,p_0) ( t 0 , p 0 ) ,
− f μ + β ε ( U − G ) = − f μ + β ε ( V μ − G ) ≥ − C ‾ f + β ε ( ε ) = − C ‾ f + β ( 1 ) ≥ 0. -f^\mu+\beta_\varepsilon(U-G)=-f^\mu+\beta_\varepsilon(V^\mu-G)\geq-\overline{C}_f+\beta_\varepsilon(\varepsilon)=-\overline{C}_f+\beta(1)\geq0. − f μ + β ε ( U − G ) = − f μ + β ε ( V μ − G ) ≥ − C f + β ε ( ε ) = − C f + β ( 1 ) ≥ 0.
Combining the last two inequalities, yields (33) .
The proof is now complete.
Proposition 3.9. Assume (22) . For every μ ∈ M \mu\in\mathcal{M} μ ∈ M and each t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , the map p ↦ V μ ( t , p ) p\mapsto V^\mu(t,p) p ↦ V μ ( t , p ) is concave on [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Proof. We prove the result by showing the claim for the solution to the auxiliary problem (32) . To deal with the additional difficulty that G G G is not and f μ f^\mu f μ may not be in C 2 C^2 C 2 , we approximate them by smooth concave functions G δ G^\delta G δ and f μ , δ f^{\mu,\delta} f μ , δ . The stability of the solutions to the penalized problem yield that the solution V μ , δ V^{\mu,\delta} V μ , δ to (32) with obstacle G δ G^\delta G δ and running cost f μ , δ f^{\mu,\delta} f μ , δ converges uniformly to the solution of the problem with obstacle G G G and running cost f μ f^\mu f μ . Hence, we only show that solutions to the regularized (32) , which for simplicity we also denote by V μ V^\mu V μ , are concave.
To ease the notation, we write w : = ∂ p 2 V μ w:=\partial_p^2V^\mu w := ∂ p 2 V μ . After differentiating the penalized equation twice with respect to p p p and using the convexity of β \beta β and the concavity of f μ f^\mu f μ and G G G , we get
− ∂ t w − a μ ∂ p 2 w − 2 ∂ p a μ ∂ p w + ( − ∂ p 2 a μ + β ε ′ ( V μ , ε − G ) ) w = ∂ p 2 f μ + β ε ′ ( V μ , ε − G ) ∂ p 2 G − β ε ′ ′ ( V μ , ε − G ) ( ∂ p V μ , ε − ∂ p G ) 2 ≤ 0. -\partial_tw-a^\mu\partial_p^2w-2\partial_pa^\mu\partial_pw+\left(-\partial_p^2a^\mu+\beta'_\varepsilon(V^{\mu,\varepsilon}-G)\right)w
= \partial_p^2f^\mu+\beta'_\varepsilon(V^{\mu,\varepsilon}-G)\partial_p^2G-\beta''_\varepsilon(V^{\mu,\varepsilon}-G)(\partial_pV^{\mu,\varepsilon}-\partial_pG)^2\leq0. − ∂ t w − a μ ∂ p 2 w − 2 ∂ p a μ ∂ p w + ( − ∂ p 2 a μ + β ε ′ ( V μ , ε − G ) ) w = ∂ p 2 f μ + β ε ′ ( V μ , ε − G ) ∂ p 2 G − β ε ′′ ( V μ , ε − G ) ( ∂ p V μ , ε − ∂ p G ) 2 ≤ 0.
Since w ( T , y ) = ∂ p 2 G ( y ) ≤ 0 w(T,y)=\partial_p^2G(y)\leq0 w ( T , y ) = ∂ p 2 G ( y ) ≤ 0 , it follows from the maximum principle that w ≤ 0 w\leq0 w ≤ 0 on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] and we conclude.
In the argument above we ignored any boundary conditions for ∂ p 2 V \partial_p^2V ∂ p 2 V without showing that they are irrelevant. There are two ways to complete the proof. The first is to show, as we did earlier in the paper, that there is no need for boundary conditions. The second is to show that w ~ = ζ w ≤ 0 \widetilde{w}=\zeta w\leq0 w = ζ w ≤ 0 , where ζ \zeta ζ is a nonnegative function that vanishes near 0 and 1. This implies that w ≤ 0 w\leq0 w ≤ 0 in the support of ζ \zeta ζ and we may conclude by letting the support of ζ \zeta ζ become ( 0 , 1 ) (0,1) ( 0 , 1 ) . We leave the details to the interested reader.
The geometry of the free boundaries
We establish some basic properties of the exploration and stopping regions E μ E^\mu E μ and S μ S^\mu S μ , which are defined in (12) and (13) respectively, and their t t t –slices are given by (43) below.
We begin by showing that the form of the obstacle G G G yields that the exploration region contains a neighborhood of the line [ 0 , T ) × { m } [0,T)\times\{m\} [ 0 , T ) × { m } .
Lemma 3.10. Assume (22) and fix μ ∈ M \mu\in\mathcal{M} μ ∈ M . Then,
V μ ( ⋅ , m ) < G ( m ) in [ 0 , T ) . (34) V^\mu(\cdot,m)<G(m)\quad\text{in }[0,T). \tag*{(34)} V μ ( ⋅ , m ) < G ( m ) in [ 0 , T ) . (34)
Proof. To prove (34), we fix t ∈ [ 0 , T ) t\in[0,T) t ∈ [ 0 , T ) , and, for ε > 0 \varepsilon>0 ε > 0 small, we choose p ± ε ∈ ( 0 , 1 ) p^\varepsilon_\pm\in(0,1) p ± ε ∈ ( 0 , 1 ) such that
G ( p ± ε ) = G ( m ) − ( C ‾ f + 1 ) ε . G(p^\varepsilon_\pm)=G(m)-(\overline{C}_f+1)\varepsilon. G ( p ± ε ) = G ( m ) − ( C f + 1 ) ε .
Let u ‾ \overline{u} u be the solution to
{ − ∂ s u ‾ − a μ ( s , p ) ∂ p 2 u ‾ − C ‾ f = 0 in [ t , t + ε ) × ( p − ε , p + ε ) , u ‾ ( ⋅ , p ± ε ) = G ( m ) − ( C ‾ f + 1 ) ε in [ t , t + ε ] , u ‾ ( t + ε , ⋅ ) = G on [ p − ε , p + ε ] . (35) \begin{cases}
-\partial_s\overline{u}-a^\mu(s,p)\partial_p^2\overline{u}-\overline{C}_f=0 & \text{in }[t,t+\varepsilon)\times(p^\varepsilon_-,p^\varepsilon_+),\\
\overline{u}(\cdot,p^\varepsilon_\pm)=G(m)-(\overline{C}_f+1)\varepsilon& \text{in }[t,t+\varepsilon],\\
\overline{u}(t+\varepsilon,\cdot)=G & \text{on }[p^\varepsilon_-,p^\varepsilon_+]. \tag*{(35)}
\end{cases} ⎩ ⎨ ⎧ − ∂ s u − a μ ( s , p ) ∂ p 2 u − C f = 0 u ( ⋅ , p ± ε ) = G ( m ) − ( C f + 1 ) ε u ( t + ε , ⋅ ) = G in [ t , t + ε ) × ( p − ε , p + ε ) , in [ t , t + ε ] , on [ p − ε , p + ε ] . (35)
Since V μ V^\mu V μ is a sub-solution to (35), it follows that
V μ ≤ u ‾ in [ t , t + ε ) × [ p − ε , p + ε ] . (36) V^\mu\leq\overline{u}\quad\text{in }[t,t+\varepsilon)\times[p^\varepsilon_-,p^\varepsilon_+]. \tag*{(36)} V μ ≤ u in [ t , t + ε ) × [ p − ε , p + ε ] . (36)
Moreover,
u ‾ ( s , p ) ≤ ( G ( m ) + C ‾ f ( t + ε − s ) ) u f ( s , p ) + ( G ( p ± ε ) + C ‾ f ( t + ε − s ) ) u b ( s , p ) \overline{u}(s,p)\leq\bigl(G(m)+\overline{C}_f(t+\varepsilon-s)\bigr)u_f(s,p)+\bigl(G(p^\varepsilon_\pm)+\overline{C}_f(t+\varepsilon-s)\bigr)u_b(s,p) u ( s , p ) ≤ ( G ( m ) + C f ( t + ε − s ) ) u f ( s , p ) + ( G ( p ± ε ) + C f ( t + ε − s ) ) u b ( s , p )
in [ t , t + ε ) × [ p − ε , p + ε ] [t,t+\varepsilon)\times[p^\varepsilon_-,p^\varepsilon_+] [ t , t + ε ) × [ p − ε , p + ε ] . {#eq:37}
where u b u_b u b and u f u_f u f solve the backward boundary value problems
− ∂ s u − a μ ( s , p ) ∂ p 2 u = 0 in [ t , t + ε ) × ( p − ε , p + ε ) . (37) -\partial_su-a^\mu(s,p)\partial_p^2u=0\quad\text{in }[t,t+\varepsilon)\times(p^\varepsilon_-,p^\varepsilon_+). \tag*{(37)} − ∂ s u − a μ ( s , p ) ∂ p 2 u = 0 in [ t , t + ε ) × ( p − ε , p + ε ) . (37)
with
u b ( ⋅ , p ± ε ) = 1 on [ t , t + ε ) and u b ( t + ε , ⋅ ) = 0 on [ p − ε , p + ε ] , u_b(\cdot,p^\varepsilon_\pm)=1\quad\text{on }[t,t+\varepsilon)\quad\text{and}\quad u_b(t+\varepsilon,\cdot)=0\quad\text{on }[p^\varepsilon_-,p^\varepsilon_+], u b ( ⋅ , p ± ε ) = 1 on [ t , t + ε ) and u b ( t + ε , ⋅ ) = 0 on [ p − ε , p + ε ] ,
and
u f ( ⋅ , p ± ε ) = 0 on [ t , t + ε ) and u f ( t + ε , ⋅ ) = 1 on [ p − ε , p + ε ] . u_f(\cdot,p^\varepsilon_\pm)=0\quad\text{on }[t,t+\varepsilon)\quad\text{and}\quad u_f(t+\varepsilon,\cdot)=1\quad\text{on }[p^\varepsilon_-,p^\varepsilon_+]. u f ( ⋅ , p ± ε ) = 0 on [ t , t + ε ) and u f ( t + ε , ⋅ ) = 1 on [ p − ε , p + ε ] .
Note that the boundary conditions and the linearity of (38) imply that
u b + u f = 1 in [ t , t + ε ] × [ p − ε , p + ε ] . (38) u_b+u_f=1\quad\text{in }[t,t+\varepsilon]\times[p^\varepsilon_-,p^\varepsilon_+]. \tag*{(38)} u b + u f = 1 in [ t , t + ε ] × [ p − ε , p + ε ] . (38)
Using (36), (37) and (39) we deduce that
V ( t , m ) − G ( m ) ≤ ε ( ( C ‾ f + 1 ) u f ( t , m ) − 1 ) . V(t,m)-G(m)\leq\varepsilon\bigl((\overline{C}_f+1)u_f(t,m)-1\bigr). V ( t , m ) − G ( m ) ≤ ε ( ( C f + 1 ) u f ( t , m ) − 1 ) .
The uniform ellipticity of (38) and elementary estimates about exit times imply that, as ε → 0 \varepsilon\to0 ε → 0 , u f ( t , m ) → 0 u_f(t,m)\to0 u f ( t , m ) → 0 , and, hence, (34) holds.
The next result establishes that the stopping region contains a, uniform in μ \mu μ , compact neighborhood of [ 0 , T ] × { 0 , 1 } [0,T]\times\{0,1\} [ 0 , T ] × { 0 , 1 } . The proof is based on the construction of uniform in μ \mu μ explicit barriers near the end points of [ 0 , 1 ] [0,1] [ 0 , 1 ] .
Lemma 3.11. Assume (22). Then, there exists κ > 0 \kappa>0 κ > 0 , which is independent of ( t , μ ) ∈ [ 0 , T ] × M (t,\mu)\in[0,T]\times\mathcal{M} ( t , μ ) ∈ [ 0 , T ] × M , such that
V μ = G in [ 0 , T ] × ( [ 0 , κ ] ∪ [ 1 − κ , 1 ] ) × M . (39) V^\mu=G \quad\text{in}\quad[0,T]\times\left([0,\kappa]\cup[1-\kappa,1]\right)\times\mathcal{M}. \tag*{(39)} V μ = G in [ 0 , T ] × ( [ 0 , κ ] ∪ [ 1 − κ , 1 ] ) × M . (39) Proof. Let
M : = c f 48 c ˉ b 2 . (40) M:=\frac{c_f}{48\bar{c}_b^2}. \tag*{(40)} M := 48 c ˉ b 2 c f . (40) and choose κ > 0 \kappa>0 κ > 0 small enough so that
3 M κ > 2 a ˉ and 2 κ < m < 1 − 2 κ , (41) \frac{3M}{\kappa}>2\bar{a}\quad\text{and}\quad2\kappa<m<1-2\kappa, \tag*{(41)} κ 3 M > 2 a ˉ and 2 κ < m < 1 − 2 κ , (41) Let
d κ : = G ( 1 − 2 κ ) − G ( 2 κ ) 1 − 4 κ and L κ ( p ) : = G ( 2 κ ) − M + d κ ( p − 2 κ ) , d_\kappa:=\frac{G(1-2\kappa)-G(2\kappa)}{1-4\kappa}
\quad\text{and}\quad
L_\kappa(p):=G(2\kappa)-M+d_\kappa(p-2\kappa), d κ := 1 − 4 κ G ( 1 − 2 κ ) − G ( 2 κ ) and L κ ( p ) := G ( 2 κ ) − M + d κ ( p − 2 κ ) , and consider the continuous function
v ( p ) : = { G ( p ) − M ( p κ − 1 ) + 3 in [ 0 , 2 κ ] , L κ ( p ) in [ 2 κ , 1 − 2 κ ] , G ( p ) − M ( 1 − p κ − 1 ) + 3 in [ 1 − 2 κ , 1 ] . v(p):=
\begin{cases}
G(p)-M\left(\dfrac{p}{\kappa}-1\right)^3_+ & \text{in }[0,2\kappa],\\
L_\kappa(p) & \text{in }[2\kappa,1-2\kappa],\\
G(p)-M\left(\dfrac{1-p}{\kappa}-1\right)^3_+ & \text{in }[1-2\kappa,1].
\end{cases} v ( p ) := ⎩ ⎨ ⎧ G ( p ) − M ( κ p − 1 ) + 3 L κ ( p ) G ( p ) − M ( κ 1 − p − 1 ) + 3 in [ 0 , 2 κ ] , in [ 2 κ , 1 − 2 κ ] , in [ 1 − 2 κ , 1 ] . Using that the chord joining ( 2 κ , G ( 2 κ ) ) (2\kappa,G(2\kappa)) ( 2 κ , G ( 2 κ )) and ( 1 − 2 κ , G ( 1 − 2 κ ) ) (1-2\kappa,G(1-2\kappa)) ( 1 − 2 κ , G ( 1 − 2 κ )) lies below G G G , we deduce that L κ ≤ G − M < G L_\kappa\leq G-M<G L κ ≤ G − M < G on [ 2 κ , 1 − 2 κ ] [2\kappa,1-2\kappa] [ 2 κ , 1 − 2 κ ] , while on the two remaining pieces, it is immediate that v ≤ G v\leq G v ≤ G .
Direct calculations yield
v ′ ′ ( p ) = G ′ ′ ( p ) − 6 M κ 2 ( p κ − 1 ) in ( κ , 2 κ ) , v''(p)=G''(p)-\frac{6M}{\kappa^2}\left(\frac{p}{\kappa}-1\right)\quad\text{in }(\kappa,2\kappa), v ′′ ( p ) = G ′′ ( p ) − κ 2 6 M ( κ p − 1 ) in ( κ , 2 κ ) , and
v ′ ′ ( p ) = G ′ ′ ( p ) − 6 M κ 2 ( 1 − p κ − 1 ) in ( 1 − 2 κ , 1 − κ ) . v''(p)=G''(p)-\frac{6M}{\kappa^2}\left(\frac{1-p}{\kappa}-1\right)\quad\text{in }(1-2\kappa,1-\kappa). v ′′ ( p ) = G ′′ ( p ) − κ 2 6 M ( κ 1 − p − 1 ) in ( 1 − 2 κ , 1 − κ ) . Since G ′ ′ = 0 G''=0 G ′′ = 0 on these intervals and a μ ( t , p ) ≤ 2 c ˉ b 2 κ 2 a^\mu(t,p)\leq2\bar{c}_b^2\kappa^2 a μ ( t , p ) ≤ 2 c ˉ b 2 κ 2 , (41) and (42) give
− a μ v ′ ′ − f μ ≤ 12 c ˉ b 2 M − c f = − 3 4 c f < 0 in ( κ , 2 κ ) ∪ ( 1 − 2 κ , 1 − κ ) . -a^\mu v''-f^\mu\leq12\bar{c}_b^2M-c_f=-\frac{3}{4}c_f<0\quad\text{in }(\kappa,2\kappa)\cup(1-2\kappa,1-\kappa). − a μ v ′′ − f μ ≤ 12 c ˉ b 2 M − c f = − 4 3 c f < 0 in ( κ , 2 κ ) ∪ ( 1 − 2 κ , 1 − κ ) . In the middle interval ( 2 κ , 1 − 2 κ ) (2\kappa,1-2\kappa) ( 2 κ , 1 − 2 κ ) , it is immediate that
− a μ v ′ ′ − f μ = − f μ ≤ − c f , -a^\mu v''-f^\mu=-f^\mu\leq-c_f, − a μ v ′′ − f μ = − f μ ≤ − c f , while, on [ 0 , κ ] ∪ [ 1 − κ , 1 ] [0,\kappa]\cup[1-\kappa,1] [ 0 , κ ] ∪ [ 1 − κ , 1 ] , we have
− a μ v ′ ′ − f μ = − f μ ≤ − c f < 0. -a^\mu v''-f^\mu=-f^\mu\leq-c_f<0. − a μ v ′′ − f μ = − f μ ≤ − c f < 0. At the two junctions, that is, at p = 2 κ p=2\kappa p = 2 κ and p = 1 − 2 κ p=1-2\kappa p = 1 − 2 κ , using (42) and the fact that ∣ d κ ∣ , ∣ G ′ ( 2 κ ) ∣ , ∣ G ′ ( 1 − 2 κ ) ∣ ≤ L G |d_\kappa|,|G'(2\kappa)|,|G'(1-2\kappa)|\leq L_G ∣ d κ ∣ , ∣ G ′ ( 2 κ ) ∣ , ∣ G ′ ( 1 − 2 κ ) ∣ ≤ L G , we find that
∂ p + v ( 2 κ ) − ∂ p − v ( 2 κ ) = d κ − G ′ ( 2 κ ) + 3 M κ > 0 and \partial_p^+v(2\kappa)-\partial_p^-v(2\kappa)=d_\kappa-G'(2\kappa)+\frac{3M}{\kappa}>0
\quad\text{and} ∂ p + v ( 2 κ ) − ∂ p − v ( 2 κ ) = d κ − G ′ ( 2 κ ) + κ 3 M > 0 and ∂ p + v ( 1 − 2 κ ) − ∂ p − v ( 1 − 2 κ ) = G ′ ( 1 − 2 κ ) + 3 M κ − d κ > 0. \partial_p^+v(1-2\kappa)-\partial_p^-v(1-2\kappa)=G'(1-2\kappa)+\frac{3M}{\kappa}-d_\kappa>0. ∂ p + v ( 1 − 2 κ ) − ∂ p − v ( 1 − 2 κ ) = G ′ ( 1 − 2 κ ) + κ 3 M − d κ > 0.
Hence, there does not exist a C 2 C^2 C 2 –test function which can touch v v v from above at either point.
It follows that v v v is a viscosity sub-solution to (10) , and, since v ≤ G v \leq G v ≤ G at t = T t=T t = T , the comparison gives
v ≤ V μ ≤ G . v \leq V^\mu\leq G. v ≤ V μ ≤ G .
Because v = G v=G v = G on the two endpoint intervals, (39) follows.
Next we recall the sets A ± μ ( t ) A^\mu_\pm(t) A ± μ ( t ) given, for each t ∈ [ 0 , T ) t\in[0,T) t ∈ [ 0 , T ) , by (14) .
The next proposition characterizes the t t t –slices E t μ \mathcal{E}^\mu_t E t μ and S t μ \mathcal{S}^\mu_t S t μ of the continuation and stopping regions in terms of A − μ ( t ) A^\mu_-(t) A − μ ( t ) and A + μ ( t ) A^\mu_+(t) A + μ ( t ) .
Proposition 3.12. Assume (22) and fix μ ∈ M \mu\in\mathcal{M} μ ∈ M . Then, there exists κ ∈ ( 0 , 1 ) \kappa\in(0,1) κ ∈ ( 0 , 1 ) such that, for every t ∈ [ 0 , T ) t\in[0,T) t ∈ [ 0 , T ) ,
0 < κ < A − μ ( t ) < m < A + μ ( t ) < 1 − κ < 1. 0<\kappa<A^\mu_-(t)<m<A^\mu_+(t)<1-\kappa<1. 0 < κ < A − μ ( t ) < m < A + μ ( t ) < 1 − κ < 1. Moreover, for each t ∈ [ 0 , T ) t\in[0,T) t ∈ [ 0 , T ) ,
E t μ = ( A − μ ( t ) , A + μ ( t ) ) and S t μ = [ 0 , A − μ ( t ) ] ∪ [ A + μ ( t ) , 1 ] . (42) \mathcal{E}^\mu_t=(A^\mu_-(t),A^\mu_+(t)) \quad\text{and}\quad\mathcal{S}^\mu_t=[0,A^\mu_-(t)]\cup[A^\mu_+(t),1]. \tag*{(42)} E t μ = ( A − μ ( t ) , A + μ ( t )) and S t μ = [ 0 , A − μ ( t )] ∪ [ A + μ ( t ) , 1 ] . (42)
Proof. For each s ∈ [ 0 , T ) s\in[0,T) s ∈ [ 0 , T ) , we use Lemma 3.11 and define the nearest contact points on the two sides of m m m by
b − μ ( s ) : = max { p ≤ m : V μ ( s , p ) = G ( p ) } and b + μ ( s ) : = min { p ≥ m : V μ ( s , p ) = G ( p ) } . b^\mu_-(s):=\max\{p\leq m:V^\mu(s,p)=G(p)\}\quad\text{and}\quad b^\mu_+(s):=\min\{p\geq m:V^\mu(s,p)=G(p)\}. b − μ ( s ) := max { p ≤ m : V μ ( s , p ) = G ( p )} and b + μ ( s ) := min { p ≥ m : V μ ( s , p ) = G ( p )} .
It follows from the definition above and Lemma 3.10 that
V μ ( s , p ) < G ( p ) if p ∈ ( b − μ ( s ) , b + μ ( s ) ) . V^\mu(s,p)<G(p)\quad\text{if}\quad p\in(b^\mu_-(s),b^\mu_+(s)). V μ ( s , p ) < G ( p ) if p ∈ ( b − μ ( s ) , b + μ ( s )) .
We prove that there is no continuation region to the right of b + μ b^\mu_+ b + μ and left of b − μ b^\mu_- b − μ , which, in turn, implies that, for all s ∈ [ 0 , T ) s\in[0,T) s ∈ [ 0 , T ) ,
A − μ ( s ) = b − μ ( s ) and A + μ ( s ) = b + μ ( s ) . A^\mu_-(s)=b^\mu_-(s)\quad\text{and}\quad A^\mu_+(s)=b^\mu_+(s). A − μ ( s ) = b − μ ( s ) and A + μ ( s ) = b + μ ( s ) .
Since the arguments are similar, we only present the details for the former.
Fix s ∈ [ 0 , T ) s\in[0,T) s ∈ [ 0 , T ) and set V ^ μ ( s , p ) : = V μ ( s , p ) − G ( p ) \widehat{V}^\mu(s,p):=V^\mu(s,p)-G(p) V μ ( s , p ) := V μ ( s , p ) − G ( p ) . It follows that V ^ μ ≤ 0 \widehat{V}^\mu\leq0 V μ ≤ 0 and V ^ μ ( s , b + μ ( s ) ) = V ^ μ ( s , 1 ) = 0 \widehat{V}^\mu(s,b^\mu_+(s))=\widehat{V}^\mu(s,1)=0 V μ ( s , b + μ ( s )) = V μ ( s , 1 ) = 0 . Moreover, in view of the linearity of G G G on [ m , 1 ] [m,1] [ m , 1 ] , Proposition 3.9 yields that the map p ↦ V ^ μ ( s , p ) p\mapsto\widehat{V}^\mu(s,p) p ↦ V μ ( s , p ) is concave on [ m , 1 ] [m,1] [ m , 1 ] .
Hence, for all p ∈ [ b + μ ( s ) , 1 ] p\in[b^\mu_+(s),1] p ∈ [ b + μ ( s ) , 1 ] ,
0 ≥ V ^ μ ( s , p ) ≥ 1 − p 1 − b + μ ( s ) V ^ μ ( s , b + μ ( s ) ) + p − b + μ ( s ) 1 − b + μ ( s ) V ^ μ ( s , 1 ) = 0. 0\geq\widehat{V}^\mu(s,p)\geq\frac{1-p}{1-b^\mu_+(s)}\widehat{V}^\mu(s,b^\mu_+(s))+\frac{p-b^\mu_+(s)}{1-b^\mu_+(s)}\widehat{V}^\mu(s,1)=0. 0 ≥ V μ ( s , p ) ≥ 1 − b + μ ( s ) 1 − p V μ ( s , b + μ ( s )) + 1 − b + μ ( s ) p − b + μ ( s ) V μ ( s , 1 ) = 0.
Applying the same argument on [ 0 , m ] [0,m] [ 0 , m ] and using that V μ ( s , 0 ) = V ^ μ ( s , b − μ ( s ) ) = 0 V^\mu(s,0)=\widehat{V}^\mu(s,b^\mu_-(s))=0 V μ ( s , 0 ) = V μ ( s , b − μ ( s )) = 0 , we deduce that
V μ ( s , p ) = G ( p ) for p ∈ [ 0 , b − μ ( s ) ] ∪ [ b + μ ( s ) , 1 ] . V^\mu(s,p)=G(p)\quad\text{for}\quad p\in[0,b^\mu_-(s)]\cup[b^\mu_+(s),1]. V μ ( s , p ) = G ( p ) for p ∈ [ 0 , b − μ ( s )] ∪ [ b + μ ( s ) , 1 ] .
This proves the required identities A ± μ ( s ) = b ± μ ( s ) A^\mu_\pm(s)=b^\mu_\pm(s) A ± μ ( s ) = b ± μ ( s ) .
The rest of the claim follows easily.
The stability of the free boundaries
We state and prove the stability of the free boundaries with respect to the weak convergence of the collective input. The argument is complicated and the analysis relies, among others, on two results obtained in [4 ] and [3 ] , which are stated below. We note that these references work with the forward in time problem for solutions to the obstacle problem with obstacle 0 0 0 , that is, they formulate their problem for G − V μ G - V^\mu G − V μ .
Since our setting, motivated by the application, was for V μ − G V^\mu- G V μ − G , below we rewrite the results of [4 ] in a form consistent with the backward in time evolution.
We begin with two important facts about the interfaces which we take from [4 ] (see also [3 ] ). To use them, we need to reformulate the problem as we explain below.
From the discussion in the previous subsection, we know that the free boundaries ( A ± μ ( t ) ) t ∈ [ 0 , T ) (A^\mu_\pm(t))_{t\in[0,T)} ( A ± μ ( t ) ) t ∈ [ 0 , T ) are in the domain [ 0 , T ) × ( [ 0 , 1 ] ∖ { m } ) [0,T) \times([0,1] \setminus\{m\}) [ 0 , T ) × ([ 0 , 1 ] ∖ { m }) .
It is a classical fact in the free boundary literature, that, around the curves ( A ± μ ( t ) ) t ∈ [ 0 , T ] (A^\mu_\pm(t))_{t\in[0,T]} ( A ± μ ( t ) ) t ∈ [ 0 , T ] we may rewrite (10) with unknown V ^ μ = V μ − G \widehat V^\mu= V^\mu- G V μ = V μ − G as
∂ t V ^ μ + a μ ∂ p 2 V ^ μ = − f μ 1 { V ^ μ < 0 } and V ^ μ ≤ 0. (43) \partial_t\widehat V^\mu+ a^\mu\partial_p^2\widehat V^\mu= -f^\mu\mathbf{1}_{\{\widehat V^\mu<0\}} \quad\text{and} \quad\widehat V^\mu\leq0. \tag*{(43)} ∂ t V μ + a μ ∂ p 2 V μ = − f μ 1 { V μ < 0 } and V μ ≤ 0. (43)
Indeed, standard parabolic estimates and passage to the penalization limit yield V ^ μ ∈ W l o c 2 , 1 ; q ( ( 0 , T ) × ( ( 0 , 1 ) ∖ { m } ) ) \widehat V^\mu\in W_{\mathrm{loc}}^{2,1;q}((0,T)\times((0,1)\setminus\{m\})) V μ ∈ W loc 2 , 1 ; q (( 0 , T ) × (( 0 , 1 ) ∖ { m })) for every 1 < q < ∞ 1<q<\infty 1 < q < ∞ . Since ∂ t V ^ μ \partial_t\widehat V^\mu ∂ t V μ and ∂ p 2 V ^ μ \partial_p^2\widehat V^\mu ∂ p 2 V μ vanish almost everywhere on { V ^ μ = 0 } \{\widehat V^\mu=0\} { V μ = 0 } , the equation in the continuation region gives (43) almost everywhere.
In what follows, to state the results we use the parabolic cylinders around a point ( s 0 , q 0 ) ∈ [ 0 , T ) × ( 0 , 1 ) (s_0,q_0)\in[0,T)\times(0,1) ( s 0 , q 0 ) ∈ [ 0 , T ) × ( 0 , 1 ) given by
Q R ( s 0 , q 0 ) = { ( s , q ) ∈ [ 0 , T ) × [ 0 , 1 ] : ∣ s − s 0 ∣ < R 2 and ∣ q − q 0 ∣ < R } , Q_R(s_0,q_0)=\{(s,q)\in[0,T)\times[0,1]: |s-s_0|<R^2\ \text{and}\ |q-q_0|<R\}, Q R ( s 0 , q 0 ) = {( s , q ) ∈ [ 0 , T ) × [ 0 , 1 ] : ∣ s − s 0 ∣ < R 2 and ∣ q − q 0 ∣ < R } ,
and
Q R + ( s 0 , q 0 ) = { ( s , q ) ∈ [ 0 , T ) × [ 0 , 1 ] : 0 ≤ s − s 0 < R 2 and ∣ q − q 0 ∣ < R } . Q_R^+(s_0,q_0)=\{(s,q)\in[0,T)\times[0,1]: 0\leq s-s_0<R^2\ \text{and}\ |q-q_0|<R\}. Q R + ( s 0 , q 0 ) = {( s , q ) ∈ [ 0 , T ) × [ 0 , 1 ] : 0 ≤ s − s 0 < R 2 and ∣ q − q 0 ∣ < R } .
We also recall that our assumptions imply that there exist some δ > 0 \delta>0 δ > 0 such that in a neighborhood of the free boundaries we have
a μ ≥ δ and f μ ≥ δ . a^\mu\geq\delta\quad\text{and}\quad f^\mu\geq\delta. a μ ≥ δ and f μ ≥ δ .
The first result is Theorem 2.1 in [4 ] , as corrected by Theorem 1 in [5 ] . We note that there is no μ \mu μ in this reference. Nevertheless, the bound depends only on the regularity and bounds on the coefficients, which, in view of (22), are uniform in μ \mu μ .
Proposition 3.13. Given ( s 0 , q 0 ) ∈ [ 0 , T ) × ( [ 0 , 1 ] ∖ { m } ) (s_0,q_0)\in[0,T)\times([0,1]\setminus\{m\}) ( s 0 , q 0 ) ∈ [ 0 , T ) × ([ 0 , 1 ] ∖ { m }) and for all Q R ( s 0 , q 0 ) ⊂ [ 0 , T ) × ( [ 0 , 1 ] ∖ { m } ) Q_R(s_0,q_0)\subset[0,T)\times([0,1]\setminus\{m\}) Q R ( s 0 , q 0 ) ⊂ [ 0 , T ) × ([ 0 , 1 ] ∖ { m }) , the V ^ μ \widehat V^\mu V μ ’s are bounded in W 2 , 1 ; ∞ ( Q R ( s 0 , q 0 ) ) W^{2,1;\infty}(Q_R(s_0,q_0)) W 2 , 1 ; ∞ ( Q R ( s 0 , q 0 )) uniformly in R ∈ ( 0 , R 0 ) R\in(0,R_0) R ∈ ( 0 , R 0 ) and μ ∈ M \mu\in\mathcal{M} μ ∈ M .
The second result we need is the following non-degeneracy lemma (Lemma 2.5 in [4 ] ), which can be interpreted as the fact that the free boundaries are not “blurred”, that is, they cannot appear or disappear suddenly. We also note that, similarly to the bounds in Proposition 3.13, the constants in the estimate below are independent of μ \mu μ .
Proposition 3.14. Let V ^ μ \widehat{V}^{\mu} V μ be a solution to (43) in Q r + ( P ) Q_r^+(P) Q r + ( P ) for P = ( s 0 , p 0 ) ∈ { V ^ μ < 0 } P=(s_0,p_0)\in\{\widehat{V}^{\mu}<0\} P = ( s 0 , p 0 ) ∈ { V μ < 0 } for sufficiently small r > 0 r>0 r > 0 . Then, there exist C ‾ \overline{C} C and r ‾ \overline{r} r , which are independent of μ ∈ M \mu\in\mathcal{M} μ ∈ M and P P P , such that, if Q r + ( P ) ∩ { V ^ μ = 0 } Q_r^+(P)\cap\{\widehat{V}^{\mu}=0\} Q r + ( P ) ∩ { V μ = 0 } is nonempty, then, for all r ∈ ( 0 , r ‾ ) r\in(0,\overline{r}) r ∈ ( 0 , r ) ,
inf ( s , q ) ∈ Q r + ( P ) V ^ μ ( s , q ) ≤ − C ‾ r 2 . \inf_{(s,q)\in Q_r^+(P)}\widehat{V}^{\mu}(s,q)\leq-\overline{C}r^2. ( s , q ) ∈ Q r + ( P ) inf V μ ( s , q ) ≤ − C r 2 .
Following [3 ] , we say that the free boundary consists of “regular” and “singular” points. The set R \mathcal{R} R of regular points has full measure and is locally a C 1 / 2 − C^{1/2-} C 1/2 − graph. Moreover, around points of R \mathcal{R} R , the free boundary is described locally by a C 1 / 2 − C^{1/2-} C 1/2 − -subgraph or by a C 1 / 2 − C^{1/2-} C 1/2 − -supergraph. Finally, if ( t , p ) (t,p) ( t , p ) is a regular point, then
lim ( s , q ) → ( t , p ) ( s , q ) ∈ { V ^ μ < 0 } ∂ t V ^ μ ( s , q ) = 0. \lim_{\substack{(s,q)\to(t,p)\\(s,q)\in\{\widehat{V}^{\mu}<0\}}}\partial_t\widehat{V}^{\mu}(s,q)=0. ( s , q ) → ( t , p ) ( s , q ) ∈ { V μ < 0 } lim ∂ t V μ ( s , q ) = 0.
We are now ready for the stability result for the free boundaries with respect to the weak convergence of measures in M \mathcal{M} M .
Theorem 3.15. Assume (22) and let μ n , μ ∈ M \mu_n,\mu\in\mathcal{M} μ n , μ ∈ M be such that, as n → ∞ n\to\infty n → ∞ , μ n ⇀ μ \mu_n\rightharpoonup\mu μ n ⇀ μ . Then, there exists a set R ⊂ ( 0 , T ) \mathcal{R}\subset(0,T) R ⊂ ( 0 , T ) of full Lebesgue measure such that, for every compact K ⊂ R K\subset\mathcal{R} K ⊂ R ,
lim n → ∞ sup t ∈ K ∣ A ± μ n ( t ) − A ± μ ( t ) ∣ = 0. \lim_{n\to\infty}\sup_{t\in K}\left|A_{\pm}^{\mu_n}(t)-A_{\pm}^{\mu}(t)\right|=0. n → ∞ lim t ∈ K sup ∣ A ± μ n ( t ) − A ± μ ( t ) ∣ = 0. In particular, for almost every t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) , A ± μ n ( t ) → A ± μ ( t ) A_{\pm}^{\mu_n}(t)\to A_{\pm}^{\mu}(t) A ± μ n ( t ) → A ± μ ( t ) .
Proof. Since the arguments are similar, we only present the details for the convergence of the A + μ n A_+^{\mu_n} A + μ n ’s to A + μ A_+^\mu A + μ .
Set V ^ n : = V ^ μ n \widehat{V}^n:=\widehat{V}^{\mu_n} V n := V μ n , V ^ : = V ^ μ \widehat{V}:=\widehat{V}^{\mu} V := V μ , A ± n : = A ± μ n A_\pm^n:=A_\pm^{\mu_n} A ± n := A ± μ n and A ± : = A ± μ A_\pm:=A_\pm^\mu A ± := A ± μ , and recall that, in view of Proposition 3.7, we know that, as n → ∞ n\to\infty n → ∞ , V ^ n → V ^ \widehat{V}^n\to\widehat{V} V n → V uniformly in [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] .
Fix t ∈ ( 0 , T ) t\in(0,T) t ∈ ( 0 , T ) and let ( t n ) n ∈ N (t_n)_{n\in\mathbb{N}} ( t n ) n ∈ N be any sequence such that, as n → ∞ n\to\infty n → ∞ , t n → t t_n\to t t n → t . Since the A + n ( t n ) A_+^n(t_n) A + n ( t n ) ’s are bounded, along a subsequence, which, for notational simplicity, we denote as the sequence, and for some A + ∗ ( t ) ∈ [ m , 1 − κ ] A_+^*(t)\in[m,1-\kappa] A + ∗ ( t ) ∈ [ m , 1 − κ ] (by Lemma 3.11), we have
A + n ( t n ) → A + ∗ ( t ) . A_+^n(t_n)\to A_+^*(t). A + n ( t n ) → A + ∗ ( t ) .
Recall that, if p ∈ ( A − ( t ) , A + ( t ) ) p\in(A_-(t),A_+(t)) p ∈ ( A − ( t ) , A + ( t )) , then V ^ ( t , p ) < 0 \widehat{V}(t,p)<0 V ( t , p ) < 0 . It then follows from the continuity of V ^ \widehat{V} V and the uniform convergence V ^ n → V ^ \widehat{V}^n\to\widehat{V} V n → V , that V ^ n ( s , p ) < 0 \widehat{V}^n(s,p)<0 V n ( s , p ) < 0 for all large n n n and all s s s sufficiently close to t t t . Proposition 3.12 then gives p ∈ ( A − n ( s ) , A + n ( s ) ) p\in(A_-^n(s),A_+^n(s)) p ∈ ( A − n ( s ) , A + n ( s )) . Letting p → A + ( t ) p\to A_+(t) p → A + ( t ) yields
A + ∗ ( t ) ≥ A + ( t ) . (44) A_+^*(t)\geq A_+(t). \tag*{(44)} A + ∗ ( t ) ≥ A + ( t ) . (44)
Next, Theorem 1.1 in [4 ] implies that there exists a set R ⊂ ( 0 , T ) \mathcal{R}\subset(0,T) R ⊂ ( 0 , T ) of full measure such that
∂ t V ^ ( s , q ) ⟶ 0 as ( s , q ) → ( t , p ) \partial_t\widehat{V}(s,q)\longrightarrow0\quad\text{as }(s,q)\to(t,p) ∂ t V ( s , q ) ⟶ 0 as ( s , q ) → ( t , p )
at every free boundary point ( t , p ) (t,p) ( t , p ) with t ∈ R t\in\mathcal{R} t ∈ R and p ∈ ( 0 , 1 ) ∖ { m } p\in(0,1)\setminus\{m\} p ∈ ( 0 , 1 ) ∖ { m } .
Fix t ∈ R t\in\mathcal{R} t ∈ R . Since A + ∗ ( t ) > m A_+^*(t)>m A + ∗ ( t ) > m by (44) , Proposition 3.14 gives C ‾ \overline{C} C and r ‾ \overline{r} r such that, for every ρ ∈ ( 0 , r ‾ ) \rho\in(0,\overline{r}) ρ ∈ ( 0 , r ) and n n n sufficiently large,
inf 0 < s − t n < ρ 2 ∣ q − A + n ( t n ) ∣ < ρ V ^ n ( s , q ) ≤ − C ‾ ρ 2 . \inf_{\substack{0<s-t_n<\rho^2\\|q-A_+^n(t_n)|<\rho}}\widehat{V}^n(s,q)\leq-\overline{C}\rho^2. 0 < s − t n < ρ 2 ∣ q − A + n ( t n ) ∣ < ρ inf V n ( s , q ) ≤ − C ρ 2 .
For each fixed ρ \rho ρ , these cylinders lie in Q 2 ρ ( t , A + ∗ ( t ) ) Q_{2\rho}(t,A^{*}_{+}(t)) Q 2 ρ ( t , A + ∗ ( t )) for all large n n n . Letting n → ∞ n \to\infty n → ∞ and using uniform convergence gives, for every sufficiently small ρ > 0 \rho> 0 ρ > 0 ,
inf Q 2 ρ ( t , A + ∗ ( t ) ) V ^ ≤ − C ‾ ρ 2 . (45) \inf_{Q_{2\rho}(t,A^{*}_{+}(t))} \widehat{V} \le-\overline{C}\rho^{2}. \tag*{(45)} Q 2 ρ ( t , A + ∗ ( t )) inf V ≤ − C ρ 2 . (45)
This, in turn, implies that
A + ∗ ( t ) = A + ( t ) . A^{*}_{+}(t)=A_{+}(t). A + ∗ ( t ) = A + ( t ) .
Indeed, if A + ∗ ( t ) > A + ( t ) A^{*}_{+}(t)>A_{+}(t) A + ∗ ( t ) > A + ( t ) , then V ^ ( t , ⋅ ) = 0 \widehat{V}(t,\cdot)=0 V ( t , ⋅ ) = 0 on a neighborhood of A + ∗ ( t ) A^{*}_{+}(t) A + ∗ ( t ) . The defining property of R \mathcal{R} R and integration in time give
sup Q r ( t , A + ∗ ( t ) ) ∣ V ^ ∣ ≤ r 2 , e s s s u p Q r ( t , A + ∗ ( t ) ) ∣ ∂ t V ^ ∣ = o ( r 2 ) . \sup_{Q_{r}(t,A^{*}_{+}(t))}|\widehat{V}| \le r^{2}, \qquad\mathop{\mathrm{ess\,sup}}_{Q_{r}(t,A^{*}_{+}(t))}|\partial_{t}\widehat{V}|=o(r^{2}). Q r ( t , A + ∗ ( t )) sup ∣ V ∣ ≤ r 2 , ess sup Q r ( t , A + ∗ ( t )) ∣ ∂ t V ∣ = o ( r 2 ) .
Taking r = 2 ρ r=2\rho r = 2 ρ contradicts (46).
The same argument for A − A_{-} A − and the arbitrariness of the subsequence give A ± μ n ( t n ) → A ± μ ( t ) A^{\mu_n}_{\pm}(t_n)\to A^{\mu}_{\pm}(t) A ± μ n ( t n ) → A ± μ ( t ) whenever t n → t ∈ R t_n\to t\in\mathcal{R} t n → t ∈ R . Taking μ n ≡ μ \mu_n\equiv\mu μ n ≡ μ also gives continuity at such t t t . Then, compactness yields uniform convergence on each compact K ⊂ R K\subset\mathcal{R} K ⊂ R .
The temporal continuity of the free boundaries
Theorem 3.15 gives the convergence of the boundaries on a full-measure set of deterministic times. However, this alone does not imply regularity at the random first-contact time, a question we address next.
The analysis in [4 ] and [3 ] as well as subsequent references is based on a careful blow up in space-time argument around each interface point and the complete characterization of the limit problems.
Here, we investigate the local continuity of the free boundaries at the contact points of the belief process that starts at P 0 P_{0} P 0 at t = 0 t=0 t = 0 . For this, we use the blow-up argument and the results of [4 ] and [3 ] to show that, at any contact point 0 < τ ( ω ) < T 0<\tau(\omega)<T 0 < τ ( ω ) < T and a.s. in ω \omega ω , the interfaces are 1 / 2 1/2 1/2 -Hölder continuous. Then, we investigate separately the continuity of the A ± μ A_{\pm}^{\mu} A ± μ 's as t → 0 t\to0 t → 0 and t → T t\to T t → T .
All these are crucial to establish the existence of the mean field equilibria in the next section, which is based on showing the continuity in probability of the best-response pair ( τ ^ 0 , p , τ ^ 0 , p d ) (\widehat{\tau}^{0,p},\widehat{\tau}^{d}_{0,p}) ( τ 0 , p , τ 0 , p d ) defined in (15).
We begin with a technical lemma which is a modification of a result of Krylov [25 ] .
Lemma 3.16. Fix μ ∈ M \mu\in\mathcal{M} μ ∈ M , and let P μ P^{\mu} P μ be the belief process (4). For every [ 0 , T ] [0,T] [ 0 , T ] -valued F μ \mathbb{F}^{\mu} F μ -stopping time σ \sigma σ and almost surely on { 0 < σ ≤ T , P σ μ ∈ ( 0 , 1 ) } \{0<\sigma\le T,P^{\mu}_{\sigma}\in(0,1)\} { 0 < σ ≤ T , P σ μ ∈ ( 0 , 1 )} , there exist C > 0 C>0 C > 0 and a sequence t k ↑ σ t_k\uparrow\sigma t k ↑ σ such that
∣ P t k μ − P σ μ ∣ ≤ C σ − t k . |P^{\mu}_{t_k}-P^{\mu}_{\sigma}|\le C\sqrt{\sigma-t_k}. ∣ P t k μ − P σ μ ∣ ≤ C σ − t k .
Proof. Since P μ − P 0 μ P^{\mu}-P^{\mu}_{0} P μ − P 0 μ is a continuous local F μ \mathbb{F}^{\mu} F μ -martingale with quadratic variation
⟨ P μ ⟩ t = ∫ 0 t ( Δ b ( s ; μ ≤ s ) ) 2 ( P s μ ) 2 ( 1 − P s μ ) 2 d s , \langle P^{\mu}\rangle_t=\int_{0}^{t}\bigl(\Delta b(s;\mu_{\le s})\bigr)^{2}(P^{\mu}_{s})^{2}(1-P^{\mu}_{s})^{2}\,\mathrm{d}s, ⟨ P μ ⟩ t = ∫ 0 t ( Δ b ( s ; μ ≤ s ) ) 2 ( P s μ ) 2 ( 1 − P s μ ) 2 d s ,
the Dambis–Dubins–Schwarz theorem [23 Section 3.4.B] provides a Brownian motion ( W ~ q ) q ≥ 0 (\widetilde W_q)_{q\geq0} ( W q ) q ≥ 0 , on, if necessary, enlarged probability space, such that, for all t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
P t μ = P 0 μ + W ~ ⟨ P μ ⟩ t P_t^\mu=P_0^\mu+\widetilde W_{\langle P^\mu\rangle_t} P t μ = P 0 μ + W ⟨ P μ ⟩ t
that is, P μ P^\mu P μ is a Brownian motion evaluated at its quadratic-variation clock and, moreover, ⟨ P μ ⟩ T ≤ c b 2 T / 16 \langle P^\mu\rangle_T\leq c_b^2T/16 ⟨ P μ ⟩ T ≤ c b 2 T /16 .
It follows from equation (2.1) of Theorem 2.1 in [25 ] that, almost surely, for every q ∈ ( 0 , c b 2 T / 16 ] q\in(0,c_b^2T/16] q ∈ ( 0 , c b 2 T /16 ] , there exist r k = 2 − j k ↓ 0 r_k=2^{-j_k}\downarrow0 r k = 2 − j k ↓ 0 with r k < q r_k<q r k < q such that, for some constant c > 0 c>0 c > 0 ,
osc [ q − r k , q ] W ~ ≤ c r k . \operatorname{osc}_{[q-r_k,q]}\widetilde W\leq c\sqrt{r_k}. osc [ q − r k , q ] W ≤ c r k .
Let q = ⟨ P μ ⟩ σ q=\langle P^\mu\rangle_\sigma q = ⟨ P μ ⟩ σ . Since P σ μ ∈ ( 0 , 1 ) P^\mu_\sigma\in(0,1) P σ μ ∈ ( 0 , 1 ) , the continuity of P μ P^\mu P μ and the bound Δ b μ ( s ) ≥ c b \Delta b^\mu(s)\geq c_b Δ b μ ( s ) ≥ c b imply that the map t ↦ ⟨ P μ ⟩ t t\mapsto\langle P^\mu\rangle_t t ↦ ⟨ P μ ⟩ t is strictly increasing on a left neighborhood of σ \sigma σ . Hence, for all sufficiently large k k k , there exist t k ↑ σ t_k\uparrow\sigma t k ↑ σ such that ⟨ P μ ⟩ t k = q − r k \langle P^\mu\rangle_{t_k}=q-r_k ⟨ P μ ⟩ t k = q − r k . It then follows that
∣ P t k − P σ ∣ ≤ c r k with r k = ⟨ P μ ⟩ σ − ⟨ P μ ⟩ t k ≤ c b 2 16 ( σ − t k ) , |P_{t_k}-P_\sigma|\leq c\sqrt{r_k}\quad\text{with}\quad r_k=\langle P^\mu\rangle_\sigma-\langle P^\mu\rangle_{t_k}\leq\frac{c_b^2}{16}(\sigma-t_k), ∣ P t k − P σ ∣ ≤ c r k with r k = ⟨ P μ ⟩ σ − ⟨ P μ ⟩ t k ≤ 16 c b 2 ( σ − t k ) ,
hence the claim.
The next result is about the temporal continuity of the free boundaries at the contact points. The proof uses the blow-up analysis in Proposition 3.2, Theorem 3.9, Proposition 4.1 and Proposition 5.5 of [4 ] , together with Proposition 2.2 and Theorem 1.1(i) of [3 ] , where we refer for some of the background.
Proposition 3.17. Assume (22), fix μ ∈ M \mu\in\mathcal{M} μ ∈ M , and let P μ : = P 0 , P 0 ; μ P^\mu:=P^{0,P_0;\mu} P μ := P 0 , P 0 ; μ and τ : = τ ^ 0 , P 0 μ \tau:=\widehat{\tau}_{0,P_0}^\mu τ := τ 0 , P 0 μ with τ ^ 0 , P 0 μ \widehat{\tau}_{0,P_0}^\mu τ 0 , P 0 μ given by (15). Then, almost surely on { 0 < τ < T } \{0<\tau<T\} { 0 < τ < T } , we have that
P τ μ = A − μ ( τ ) or P τ μ = A + μ ( τ ) . P_\tau^\mu=A_-^\mu(\tau)\quad\text{or}\quad P_\tau^\mu=A_+^\mu(\tau). P τ μ = A − μ ( τ ) or P τ μ = A + μ ( τ ) . Moreover, there exist constants C , r > 0 C,r>0 C , r > 0 , which may depend on μ \mu μ and the realized contact point, such that, for all s , s ′ ∈ ( τ − r , τ + r ) ∩ ( 0 , T ) s,s'\in(\tau-r,\tau+r)\cap(0,T) s , s ′ ∈ ( τ − r , τ + r ) ∩ ( 0 , T ) , if P τ μ = A − μ ( τ ) P_\tau^\mu=A_-^\mu(\tau) P τ μ = A − μ ( τ ) , then
∣ A − μ ( s ) − A − μ ( s ′ ) ∣ ≤ C ∣ s − s ′ ∣ , |A_-^\mu(s)-A_-^\mu(s')|\leq C\sqrt{|s-s'|}, ∣ A − μ ( s ) − A − μ ( s ′ ) ∣ ≤ C ∣ s − s ′ ∣ , and, if P τ μ = A + μ ( τ ) P_\tau^\mu=A_+^\mu(\tau) P τ μ = A + μ ( τ ) , then
∣ A + μ ( s ) − A + μ ( s ′ ) ∣ ≤ C ∣ s − s ′ ∣ . |A_+^\mu(s)-A_+^\mu(s')|\leq C\sqrt{|s-s'|}. ∣ A + μ ( s ) − A + μ ( s ′ ) ∣ ≤ C ∣ s − s ′ ∣ .
Proof. Let Ω 0 \Omega_0 Ω 0 be a probability-one event carrying the conclusions of Lemma 3.16 with σ = τ \sigma=\tau σ = τ and the continuity of the posterior process. In what follows, we work pathwise for each fixed ω ˉ ∈ Ω 0 ∩ { 0 < τ < T } \bar\omega\in\Omega_0\cap\{0<\tau<T\} ω ˉ ∈ Ω 0 ∩ { 0 < τ < T } .
We set p 0 : = P τ μ ∈ S τ μ ∩ ( 0 , 1 ) p_0:=P_\tau^\mu\in S_\tau^\mu\cap(0,1) p 0 := P τ μ ∈ S τ μ ∩ ( 0 , 1 ) . Proposition 3.12 then implies that ( τ , p 0 ) (\tau,p_0) ( τ , p 0 ) is a space-time free-boundary point with p 0 ≠ m p_0\ne m p 0 = m .
We recall that, as discussed earlier, near the free boundaries, V ^ μ \widehat V^\mu V μ solves, in a small parabolic cubes Q R Q_R Q R containing ( τ , p 0 ) (\tau,p_0) ( τ , p 0 ) , the problem
∂ t V ^ μ + a μ ∂ p 2 V ^ μ = − f μ 1 { V ^ μ < 0 } a.e. on Q R , (46) \partial_t\widehat V^\mu+a^\mu\partial_p^2\widehat V^\mu=-f^\mu\mathbf{1}_{\{\widehat V^\mu<0\}}\quad\text{a.e. on }Q_R, \tag*{(46)} ∂ t V μ + a μ ∂ p 2 V μ = − f μ 1 { V μ < 0 } a.e. on Q R , (46)
with a μ a^\mu a μ and f μ f^\mu f μ bounded from below away from zero.
Choose R 0 > 0 R_0>0 R 0 > 0 with Q 2 R 0 ( τ , p 0 ) ⊂ ( 0 , T ) × ( ( 0 , 1 ) ∖ { m } ) Q_{2R_0}(\tau,p_0)\subset(0,T)\times((0,1)\setminus\{m\}) Q 2 R 0 ( τ , p 0 ) ⊂ ( 0 , T ) × (( 0 , 1 ) ∖ { m }) . Proposition 3.13 gives a constant C R 0 > 0 C_{R_0}>0 C R 0 > 0 such that
∥ V ^ μ ∥ W 2 , 1 ; ∞ ( Q R 0 ( τ , p 0 ) ) ≤ C R 0 . (47) \|\widehat V^\mu\|_{W^{2,1;\infty}(Q_{R_0}(\tau,p_0))}\le C_{R_0}. \tag*{(47)} ∥ V μ ∥ W 2 , 1 ; ∞ ( Q R 0 ( τ , p 0 )) ≤ C R 0 . (47)
Following [4 ] and [3 ] , we perform a blow-up argument near the point ( τ , p 0 ) (\tau,p_0) ( τ , p 0 ) . We set a 0 μ : = a μ ( τ , p 0 ) a^\mu_0:=a^\mu(\tau,p_0) a 0 μ := a μ ( τ , p 0 ) and f 0 μ : = f μ ( τ , p 0 ) > 0 f^\mu_0:=f^\mu(\tau,p_0)>0 f 0 μ := f μ ( τ , p 0 ) > 0 , and consider the family
V ^ μ , ε ( s , p ) : = 1 ε 2 V ^ μ ( τ + ε 2 f 0 μ s , p 0 + ε a 0 μ f 0 μ p ) . \widehat V^{\mu,\varepsilon}(s,p):=\frac{1}{\varepsilon^2}\widehat V^\mu\left(\tau+\frac{\varepsilon^2}{f^\mu_0}s,p_0+\varepsilon\sqrt{\frac{a^\mu_0}{f^\mu_0}}p\right). V μ , ε ( s , p ) := ε 2 1 V μ ( τ + f 0 μ ε 2 s , p 0 + ε f 0 μ a 0 μ p ) .
In view of (47) and [4 Proposition 3.2] , every sequence ε → 0 \varepsilon\to0 ε → 0 admits a subsequence along which the V ^ μ , ε \widehat V^{\mu,\varepsilon} V μ , ε ’s converge, locally uniformly in R × R \mathbb{R}\times\mathbb{R} R × R , to a solution V ^ μ , 0 : R × R → ( − ∞ , 0 ] \widehat V^{\mu,0}:\mathbb{R}\times\mathbb{R}\to(-\infty,0] V μ , 0 : R × R → ( − ∞ , 0 ] to
∂ s V ^ μ , 0 + ∂ p 2 V ^ μ , 0 = − 1 { V ^ μ , 0 < 0 } a.e. in R × R , \partial_s\widehat V^{\mu,0}+\partial_p^2\widehat V^{\mu,0}=-\mathbf{1}_{\{\widehat V^{\mu,0}<0\}}\quad\text{a.e. in }\mathbb{R}\times\mathbb{R}, ∂ s V μ , 0 + ∂ p 2 V μ , 0 = − 1 { V μ , 0 < 0 } a.e. in R × R ,
and
( 0 , 0 ) ∈ ∂ { V ^ μ , 0 = 0 } . (0,0)\in\partial\{\widehat V^{\mu,0}=0\}. ( 0 , 0 ) ∈ ∂ { V μ , 0 = 0 } .
These blow-up limits are classified by Proposition 3.8, Theorem 3.9, and Proposition 4.1 in [4 ] , after changing sign and reversing time.
It follows that all possible limits have either one of the “regular” profiles
V ^ μ , 0 ( s , p ) = − 1 2 ( max { p , 0 } ) 2 or V ^ μ , 0 ( s , p ) = − 1 2 ( max { − p , 0 } ) 2 , (48) \widehat V^{\mu,0}(s,p)=-\frac{1}{2}(\max\{p,0\})^2\quad\text{or}\quad\widehat V^{\mu,0}(s,p)=-\frac{1}{2}(\max\{-p,0\})^2, \tag*{(48)} V μ , 0 ( s , p ) = − 2 1 ( max { p , 0 } ) 2 or V μ , 0 ( s , p ) = − 2 1 ( max { − p , 0 } ) 2 , (48)
or are singular, in which case, for some c ∈ [ 0 , 1 / 2 ] c\in[0,1/2] c ∈ [ 0 , 1/2 ] ,
V ^ μ , 0 ( 0 , p ) = − c p 2 . \widehat V^{\mu,0}(0,p)=-cp^2. V μ , 0 ( 0 , p ) = − c p 2 .
Since p 0 ∈ S τ μ p_0\in S^\mu_\tau p 0 ∈ S τ μ , either p 0 ≤ A − μ ( τ ) p_0\le A^\mu_-(\tau) p 0 ≤ A − μ ( τ ) or p 0 ≥ A + μ ( τ ) p_0\ge A^\mu_+(\tau) p 0 ≥ A + μ ( τ ) . If p 0 ≤ A − μ ( τ ) p_0\le A^\mu_-(\tau) p 0 ≤ A − μ ( τ ) , for each fixed p ≤ 0 p\le0 p ≤ 0 and ε \varepsilon ε small enough, we have p 0 + ε a 0 μ f 0 μ p ≤ p 0 ≤ A − μ ( τ ) p_0+\varepsilon\sqrt{\frac{a^\mu_0}{f^\mu_0}}p\le p_0\le A^\mu_-(\tau) p 0 + ε f 0 μ a 0 μ p ≤ p 0 ≤ A − μ ( τ ) and, hence,
V ^ μ , ε ( 0 , p ) = 1 ε 2 V ^ μ ( τ , p 0 + ε a 0 μ f 0 μ p ) = 0 , \widehat V^{\mu,\varepsilon}(0,p)=\frac{1}{\varepsilon^2}\widehat V^\mu\left(\tau,p_0+\varepsilon\sqrt{\frac{a^\mu_0}{f^\mu_0}}p\right)=0, V μ , ε ( 0 , p ) = ε 2 1 V μ ( τ , p 0 + ε f 0 μ a 0 μ p ) = 0 ,
which implies that
V ^ μ , 0 ( 0 , p ) = 0 if p ≤ 0. \widehat V^{\mu,0}(0,p)=0\quad\text{if}\quad p\le0. V μ , 0 ( 0 , p ) = 0 if p ≤ 0.
If p 0 ≥ A + μ ( τ ) p_0\ge A^\mu_+(\tau) p 0 ≥ A + μ ( τ ) , for each fixed p ≥ 0 p\ge0 p ≥ 0 and ε \varepsilon ε small enough, we have p 0 + ε a 0 μ f 0 μ p ≥ p 0 ≥ A + μ ( τ ) p_0+\varepsilon\sqrt{\frac{a^\mu_0}{f^\mu_0}}p\ge p_0\ge A^\mu_+(\tau) p 0 + ε f 0 μ a 0 μ p ≥ p 0 ≥ A + μ ( τ ) and, thus,
V ^ μ , ε ( 0 , p ) = 1 ε 2 V ^ μ ( τ , p 0 + ε a 0 μ f 0 μ p ) = 0 and V ^ μ , 0 ( 0 , p ) = 0 if p ≥ 0. \widehat V^{\mu,\varepsilon}(0,p)=\frac{1}{\varepsilon^2}\widehat V^\mu\left(\tau,p_0+\varepsilon\sqrt{\frac{a^\mu_0}{f^\mu_0}}p\right)=0\quad\text{and}\quad\widehat V^{\mu,0}(0,p)=0\quad\text{if}\quad p\ge0. V μ , ε ( 0 , p ) = ε 2 1 V μ ( τ , p 0 + ε f 0 μ a 0 μ p ) = 0 and V μ , 0 ( 0 , p ) = 0 if p ≥ 0.
As a result, in the singular case we must have c = 0 c=0 c = 0 , so the classification leaves only the temporal solution
V ^ μ , 0 ( s , p ) = min { − s , 0 } = − s + . \widehat V^{\mu,0}(s,p)=\min\{-s,0\}=-s^+. V μ , 0 ( s , p ) = min { − s , 0 } = − s + .
We now rule out this limit. At a singular contact point, Proposition 5.5 in [4 ] gives the same limit along every blow-up sequence. Indeed, applying Lemma 3.16 with σ = τ \sigma=\tau σ = τ and using the definition of τ \tau τ , we can find t k ↑ τ t_k \uparrow\tau t k ↑ τ such that
∣ P t k μ − p 0 ∣ ≤ C τ − t k and V ^ μ ( t k , P t k μ ) < 0. (49) \left|P^\mu_{t_k}-p_0\right|\le C\sqrt{\tau-t_k}\quad\text{and}\quad\widehat V^\mu(t_k,P^\mu_{t_k})<0. \tag*{(49)} P t k μ − p 0 ≤ C τ − t k and V μ ( t k , P t k μ ) < 0. (49)
Set
ε k 2 : = f 0 ( τ − t k ) and p k : = P μ ( t k ) − p 0 a 0 μ ( τ − t k ) . \varepsilon_k^2:=f_0(\tau-t_k)\quad\text{and}\quad p_k:=\frac{P^\mu(t_k)-p_0}{\sqrt{a^\mu_0(\tau-t_k)}}. ε k 2 := f 0 ( τ − t k ) and p k := a 0 μ ( τ − t k ) P μ ( t k ) − p 0 .
Then, in view of (49) , we know that the p k p_k p k ’s are bounded and, after passing, if necessary, to a subsequence, we may assume that p k → p p_k\to p p k → p , while we have V ^ μ , ε k → V ^ μ , 0 \widehat V^{\mu,\varepsilon_k}\to\widehat V^{\mu,0} V μ , ε k → V μ , 0 locally uniformly and
V ^ μ , ε k ( − 1 , p k ) < 0. (50) \widehat V^{\mu,\varepsilon_k}(-1,p_k)<0. \tag*{(50)} V μ , ε k ( − 1 , p k ) < 0. (50)
The rescaled coefficients have uniform local Hölder and positive lower bounds, so the nondegeneracy constants are uniform in all sufficiently large k k k .
Choose δ ∈ ( 0 , r ‾ ∧ 1 / 2 ) \delta\in(0,\overline r\wedge1/2) δ ∈ ( 0 , r ∧ 1/2 ) , with r ‾ \overline r r given in Proposition 3.14 for the rescaled equations, and set
D : = { ( s , y ) : ∣ y − p ∣ < δ 2 , 0 < s + 1 < δ 2 4 } . D:=\left\{(s,y):|y-p|<\frac{\delta}{2},\ 0<s+1<\frac{\delta^2}{4}\right\}. D := { ( s , y ) : ∣ y − p ∣ < 2 δ , 0 < s + 1 < 4 δ 2 } .
Since V ^ μ , 0 ( s , y ) = − s + = 0 \widehat V^{\mu,0}(s,y)=-s^+=0 V μ , 0 ( s , y ) = − s + = 0 in D D D , Proposition 3.3 in [4 ] and dominated convergence give that, as k → ∞ k\to\infty k → ∞ ,
∣ D ∩ { V ^ μ , ε k < 0 } ∣ = ∫ D 1 { V ^ μ , ε k < 0 } ( s , q ) d q d s ⟶ 0. \left|D\cap\{\widehat V^{\mu,\varepsilon_k}<0\}\right|=\int_D\mathbf{1}_{\{\widehat V^{\mu,\varepsilon_k}<0\}}(s,q)\,dq\,ds\longrightarrow0. D ∩ { V μ , ε k < 0 } = ∫ D 1 { V μ , ε k < 0 } ( s , q ) d q d s ⟶ 0.
In view of the facts that ∣ D ∣ > 0 |D|>0 ∣ D ∣ > 0 and V ^ μ , ε k ≤ 0 \widehat V^{\mu,\varepsilon_k}\le0 V μ , ε k ≤ 0 , we then deduce that, for all sufficiently large k k k ,
D ∩ { V ^ μ , ε k = 0 } ≠ ∅ . D\cap\{\widehat V^{\mu,\varepsilon_k}=0\}\ne\varnothing. D ∩ { V μ , ε k = 0 } = ∅ .
Moreover, since p k → p p_k\to p p k → p , for all sufficiently large k k k , every point ( s , y ) ∈ D (s,y)\in D ( s , y ) ∈ D satisfies ∣ y − p k ∣ < δ |y-p_k|<\delta ∣ y − p k ∣ < δ and 0 < s + 1 < δ 2 0<s+1<\delta^2 0 < s + 1 < δ 2 , and, thus, V ^ μ , ε k \widehat V^{\mu,\varepsilon_k} V μ , ε k has a zero in the forward cylinder { ∣ y − p k ∣ < δ , − 1 < s < − 1 + δ 2 } \{|y-p_k|<\delta,\ -1<s<-1+\delta^2\} { ∣ y − p k ∣ < δ , − 1 < s < − 1 + δ 2 } .
It follows from (50) and Proposition 3.14 (see [3 Lemma 2.1] ) that there exists a constant C ‾ > 0 \overline C>0 C > 0 , which is independent of all sufficiently large k k k , such that, for this fixed δ \delta δ ,
inf ∣ y − p k ∣ < δ − 1 < s < − 1 + δ 2 V ^ μ , ε k ( s , y ) ≤ − C ‾ δ 2 . \inf_{\substack{|y-p_k|<\delta\\-1<s<-1+\delta^2}}\widehat V^{\mu,\varepsilon_k}(s,y)\le-\overline C\delta^2. ∣ y − p k ∣ < δ − 1 < s < − 1 + δ 2 inf V μ , ε k ( s , y ) ≤ − C δ 2 .
Because p k → p p_k\to p p k → p and δ < 1 / 2 \delta<1/2 δ < 1/2 , these cylinders lie in a fixed compact subset of { s < 0 } \{s<0\} { s < 0 } , where V ^ μ , 0 = 0 \widehat V^{\mu,0}=0 V μ , 0 = 0 . Hence, the local uniform convergence V ^ μ , ε k → V ^ μ , 0 \widehat V^{\mu,\varepsilon_k}\to\widehat V^{\mu,0} V μ , ε k → V μ , 0 gives
sup ∣ y − p k ∣ < δ − 1 < s < − 1 + δ 2 ∣ V ^ μ , ε k ( s , y ) ∣ ⟶ 0 , \sup_{\substack{|y-p_k|<\delta\\-1<s<-1+\delta^2}}\left|\widehat V^{\mu,\varepsilon_k}(s,y)\right|\longrightarrow0, ∣ y − p k ∣ < δ − 1 < s < − 1 + δ 2 sup V μ , ε k ( s , y ) ⟶ 0 ,
which is a contradiction.
In conclusion, only the two regular limits in (48) remain. But then, the strict inequality p 0 < A − μ ( τ ) p_0<A^\mu_-(\tau) p 0 < A − μ ( τ ) or p 0 > A + μ ( τ ) p_0>A^\mu_+(\tau) p 0 > A + μ ( τ ) would force an identically zero spatial trace, which contradicts these profiles. Hence p 0 = A − μ ( τ ) p_0=A^\mu_-(\tau) p 0 = A − μ ( τ ) or p 0 = A + μ ( τ ) p_0=A^\mu_+(\tau) p 0 = A + μ ( τ ) . Then, in view of [3 Proposition 2.2] , ( τ , p 0 ) (\tau,p_0) ( τ , p 0 ) is a regular point.
Next, we apply Theorem 1.1(i) in [3 ] to (46) , the hypotheses of which are satisfied. Indeed, our assumptions guarantee that the diffusion coefficients and f μ f^\mu f μ in (46) are Hölder continuous and bounded below away from 0 on the cylinder Q R 0 ( τ , p 0 ) Q_{R_0}(\tau,p_0) Q R 0 ( τ , p 0 ) used in (47) , and (47) gives V ^ μ ∈ W l o c 2 , 1 ; ∞ \widehat{V}^\mu\in W^{2,1;\infty}_{\mathrm{loc}} V μ ∈ W loc 2 , 1 ; ∞ .
It follows that, since ( τ , p 0 ) (\tau,p_0) ( τ , p 0 ) is a regular point, there exist ρ ∈ ( 0 , R 0 ) \rho\in(0,R_0) ρ ∈ ( 0 , R 0 ) and a function
g ∈ C 1 / 2 ( ( τ − ρ 2 , τ + ρ 2 ) ) , g ( τ ) = p 0 , g \in C^{1/2}\left((\tau-\rho^2,\tau+\rho^2)\right), \qquad g(\tau)=p_0, g ∈ C 1/2 ( ( τ − ρ 2 , τ + ρ 2 ) ) , g ( τ ) = p 0 ,
such that, in Q ρ ( τ , p 0 ) ⋐ ( 0 , T ) × ( ( 0 , 1 ) ∖ { m } ) Q_\rho(\tau,p_0) \Subset(0,T) \times((0,1)\setminus\{m\}) Q ρ ( τ , p 0 ) ⋐ ( 0 , T ) × (( 0 , 1 ) ∖ { m }) , the stopping set { V ^ μ = 0 } \{\widehat{V}^\mu=0\} { V μ = 0 } is either { p ≤ g ( s ) } \{p\le g(s)\} { p ≤ g ( s )} or { p ≥ g ( s ) } \{p\ge g(s)\} { p ≥ g ( s )} . Finally, we use the continuity of g g g to choose 0 < r < ρ 2 0<r<\rho^2 0 < r < ρ 2 so that ∣ g ( s ) − p 0 ∣ < ρ |g(s)-p_0|<\rho ∣ g ( s ) − p 0 ∣ < ρ whenever ∣ s − τ ∣ < r |s-\tau|<r ∣ s − τ ∣ < r .
Next, we identify g g g as A − μ A^\mu_- A − μ or A + μ A^\mu_+ A + μ , depending on which boundary is hit.
Next, suppose that p 0 = A − μ ( τ ) p_0=A^\mu_-(\tau) p 0 = A − μ ( τ ) . Then the cylinder Q ρ ( τ , p 0 ) Q_\rho(\tau,p_0) Q ρ ( τ , p 0 ) lies below m m m . In view of Proposition 3.12, at time s = τ s=\tau s = τ and for ∣ p − p 0 ∣ < ρ |p-p_0|<\rho ∣ p − p 0 ∣ < ρ , stopping is optimal exactly when p ≤ p 0 = A − μ ( τ ) p\le p_0=A^\mu_-(\tau) p ≤ p 0 = A − μ ( τ ) . Hence, in this neighborhood,
V ^ μ ( s , p ) = 0 ⟺ p ≤ g ( s ) , ∣ s − τ ∣ < r , ∣ p − p 0 ∣ < ρ . \widehat{V}^\mu(s,p)=0 \quad\Longleftrightarrow\quad p\le g(s), \qquad|s-\tau|<r,\quad|p-p_0|<\rho. V μ ( s , p ) = 0 ⟺ p ≤ g ( s ) , ∣ s − τ ∣ < r , ∣ p − p 0 ∣ < ρ .
Since g ( s ) g(s) g ( s ) is a stopping point smaller than m m m , g ( s ) ≤ A − μ ( s ) g(s)\le A^\mu_-(s) g ( s ) ≤ A − μ ( s ) . Then, every p ∈ ( g ( s ) , p 0 + ρ ) p\in(g(s),p_0+\rho) p ∈ ( g ( s ) , p 0 + ρ ) is a continuation point, so A − μ ( s ) < p A^\mu_-(s)<p A − μ ( s ) < p . Letting p → g ( s ) p\to g(s) p → g ( s ) gives A − μ ( s ) ≤ g ( s ) A^\mu_-(s)\le g(s) A − μ ( s ) ≤ g ( s ) . Therefore, for s ∈ ( τ − r , τ + r ) s\in(\tau-r,\tau+r) s ∈ ( τ − r , τ + r ) ,
A − μ ( s ) = g ( s ) , A^\mu_-(s)=g(s), A − μ ( s ) = g ( s ) ,
and, thus,
∣ A − μ ( s ) − A − μ ( s ′ ) ∣ ≤ [ g ] C 1 / 2 ∣ s − s ′ ∣ 1 / 2 . |A^\mu_-(s)-A^\mu_-(s')|\le[g]_{C^{1/2}}|s-s'|^{1/2}. ∣ A − μ ( s ) − A − μ ( s ′ ) ∣ ≤ [ g ] C 1/2 ∣ s − s ′ ∣ 1/2 .
The case p 0 = A + μ ( τ ) p_0=A^\mu_+(\tau) p 0 = A + μ ( τ ) follows similarly, yielding A + μ ( s ) = g ( s ) A^\mu_+(s)=g(s) A + μ ( s ) = g ( s ) for s ∈ ( τ − r , τ + r ) s\in(\tau-r,\tau+r) s ∈ ( τ − r , τ + r ) and the same Hölder estimate.
Taking C = 1 + [ g ] C 1 / 2 ( ( τ − r , τ + r ) ) C=1+[g]_{C^{1/2}((\tau-r,\tau+r))} C = 1 + [ g ] C 1/2 (( τ − r , τ + r )) , with r r r chosen above, we conclude the proof.
Remark 3.18. Blanchet’s theorem in [3 ] does not show that the entire free boundary is regular. The singular profile V ^ μ , 0 ( s , p ) = − s + \widehat{V}^{\mu,0}(s,p)=-s^+ V μ , 0 ( s , p ) = − s + may occur and the theorem gives local C 0 , 1 / 2 C^{0,1/2} C 0 , 1/2 -regularity only at points already known to be regular. The first contact point has an additional property which does not hold at an arbitrary free boundary point. Indeed, Lemma 3.16 provides continuation points approaching the free boundary point at a parabolic scale, which become, after rescaling, bounded points ( − 1 , p k ) (-1,p_k) ( − 1 , p k ) satisfying V ^ μ , ε k ( − 1 , p k ) < 0 \widehat{V}^{\mu,\varepsilon_k}(-1,p_k)<0 V μ , ε k ( − 1 , p k ) < 0 . This is, however, incompatible with − s + -s^+ − s + , which vanishes for s < 0 s<0 s < 0 . Since the two boundary geometry excludes the remaining singular profiles, the first contact point is regular, and Blanchet’s theorem then yields the local C 1 / 2 C^{1/2} C 1/2 boundary graph.
Next, we turn to the continuity at t = 0 t=0 t = 0 which is necessary to examine what happens when τ = 0 \tau=0 τ = 0 .
Proposition 3.19. For every μ ∈ M \mu\in\mathcal{M} μ ∈ M ,
lim t → 0 A ± μ ( t ) = A ± μ ( 0 ) . \lim_{t\to0} A^\mu_\pm(t)=A^\mu_\pm(0). t → 0 lim A ± μ ( t ) = A ± μ ( 0 ) .
Proof. The proofs are similar for A − μ A^\mu_- A − μ and A + μ A^\mu_+ A + μ , we only show the details for the former.
Suppose that t n → 0 t_n\to0 t n → 0 and A − μ ( t n ) → a A^\mu_-(t_n)\to a A − μ ( t n ) → a . Since it is immediate that a ≤ A − μ ( 0 ) a\le A^\mu_-(0) a ≤ A − μ ( 0 ) , we assume that a < A − μ ( 0 ) a<A^\mu_-(0) a < A − μ ( 0 ) . Then, for every z ∈ ( a , A − μ ( 0 ) ) z \in(a,A^\mu_-(0)) z ∈ ( a , A − μ ( 0 )) , we know that V ^ μ ( 0 , ⋅ ) = 0 \widehat{V}^\mu(0,\cdot)=0 V μ ( 0 , ⋅ ) = 0 in a neighborhood I z I_z I z in [ 0 , 1 ] [0,1] [ 0 , 1 ] of z z z , while V ^ μ ( t n , z ) < 0 \widehat{V}^\mu(t_n,z)<0 V μ ( t n , z ) < 0 for all large n n n .
Moreover, since, for some small h > 0 h>0 h > 0 ,
∂ t V ^ μ + a μ ∂ p 2 V ^ μ = − f μ in [ 0 , h ) × I z , \partial_t\widehat{V}^\mu+a\mu\partial_p^2\widehat{V}^\mu=-f^\mu\quad\text{in }[0,h)\times I_z, ∂ t V μ + a μ ∂ p 2 V μ = − f μ in [ 0 , h ) × I z ,
it follows that
∂ t + V ^ μ ( 0 , p ) = − f μ in I z . (51) \partial_t^+\widehat{V}^\mu(0,p)=-f^\mu\quad\text{in }I_z. \tag*{(51)} ∂ t + V μ ( 0 , p ) = − f μ in I z . (51)
where
∂ t + V ^ μ ( 0 , p ) = lim t → 0 + V ^ μ ( t , p ) t . \partial_t^+\widehat{V}^\mu(0,p)=\lim_{t\to0+}\frac{\widehat{V}^\mu(t,p)}{t}. ∂ t + V μ ( 0 , p ) = t → 0 + lim t V μ ( t , p ) .
Now choose a < p < q < A − μ ( 0 ) a<p<q<A^\mu_-(0) a < p < q < A − μ ( 0 ) . Then, the concavity of the map p ↦ V ^ μ ( t n , p ) p\mapsto\widehat{V}^\mu(t_n,p) p ↦ V μ ( t n , p ) and the fact that V ^ μ ( t n , A − μ ( t n ) ) = 0 \widehat{V}^\mu(t_n,A^\mu_-(t_n))=0 V μ ( t n , A − μ ( t n )) = 0 give
V ^ μ ( t n , p ) p − A − μ ( t n ) ≥ V ^ μ ( t n , q ) q − A − μ ( t n ) . \frac{\widehat{V}^\mu(t_n,p)}{p-A^\mu_-(t_n)}
\ge
\frac{\widehat{V}^\mu(t_n,q)}{q-A^\mu_-(t_n)}. p − A − μ ( t n ) V μ ( t n , p ) ≥ q − A − μ ( t n ) V μ ( t n , q ) .
Using (52), dividing by t n t_n t n , and letting n → ∞ n\to\infty n → ∞ yields
f μ ( 0 , p ) p − a ≤ f μ ( 0 , q ) q − a , \frac{f^\mu(0,p)}{p-a}\le\frac{f^\mu(0,q)}{q-a}, p − a f μ ( 0 , p ) ≤ q − a f μ ( 0 , q ) ,
a contradiction to the concavity and strict positivity of f μ ( 0 , ⋅ ) f^\mu(0,\cdot) f μ ( 0 , ⋅ ) , since
f μ ( 0 , p ) ≥ q − p q − a f μ ( 0 , a ) + p − a q − a f μ ( 0 , q ) > p − a q − a f μ ( 0 , q ) , f^\mu(0,p)\ge\frac{q-p}{q-a}f^\mu(0,a)+\frac{p-a}{q-a}f^\mu(0,q)>\frac{p-a}{q-a}f^\mu(0,q), f μ ( 0 , p ) ≥ q − a q − p f μ ( 0 , a ) + q − a p − a f μ ( 0 , q ) > q − a p − a f μ ( 0 , q ) ,
Therefore, it must be that lim inf t ↓ 0 A − μ ( t ) ≥ A − μ ( 0 ) \liminf_{t\downarrow0}A^\mu_-(t)\ge A^\mu_-(0) lim inf t ↓ 0 A − μ ( t ) ≥ A − μ ( 0 ) .
The reverse upper bound follows from the continuity of V ^ μ \widehat{V}^\mu V μ and Proposition 3.12, completing the proof for A − μ A^\mu_- A − μ .
To complete the results for the temporal continuity properties of the free boundaries, we investigate next their behavior as t → T t\to T t → T . Since V μ ( T , ⋅ ) = G V^\mu(T,\cdot)=G V μ ( T , ⋅ ) = G , there is no interface at t = T t=T t = T . Nevertheless, as it is shown next, there is a trace at p = m p=m p = m . Thus, defining A − μ ( T ) = A + μ ( T ) = m A^\mu_-(T)=A^\mu_+(T)=m A − μ ( T ) = A + μ ( T ) = m gives a continuous extension at T T T .
Proposition 3.20. Assume (22). Then, as t → T t\to T t → T and uniformly in μ ∈ M \mu\in\mathcal{M} μ ∈ M ,
lim sup t → T max { m − A − μ ( t ) , A + μ ( t ) − m } = 0. \limsup_{t\to T}\max\{m-A^\mu_-(t),A^\mu_+(t)-m\}=0. t → T lim sup max { m − A − μ ( t ) , A + μ ( t ) − m } = 0.
Proof. Fix μ ∈ M \mu\in\mathcal{M} μ ∈ M , t < T t<T t < T , p ≠ m p\ne m p = m , write P = P t , p ; μ P=P^{t,p;\mu} P = P t , p ; μ , recall that the volatility of P P P is bounded from above by c ˉ b / 4 \bar c_b/4 c ˉ b /4 , and, for any τ ∈ T t , T μ \tau\in\mathcal{T}^\mu_{t,T} τ ∈ T t , T μ , let H = inf { s ∈ [ t , T ] : P s = m } H=\inf\{s\in[t,T]:P_s=m\} H = inf { s ∈ [ t , T ] : P s = m } , with inf ∅ = ∞ \inf\varnothing=\infty inf ∅ = ∞ . Then, Itô’s formula applied to τ \tau τ yields
∣ p − m ∣ 2 P ( H ≤ τ ) ≤ E ∣ P τ ∧ H − p ∣ 2 ≤ c ˉ b 2 16 E ( τ − t ) . |p-m|^2\mathbb{P}(H\le\tau)\le\mathbb{E}|P_{\tau\wedge H}-p|^2\le\frac{\bar c_b^2}{16}\mathbb{E}(\tau-t). ∣ p − m ∣ 2 P ( H ≤ τ ) ≤ E ∣ P τ ∧ H − p ∣ 2 ≤ 16 c ˉ b 2 E ( τ − t ) .
Since G G G is linear on each side of m m m , the optional sampling theorem gives E G ( P τ ∧ H ) = G ( p ) \mathbb{E}G(P_{\tau\wedge H})=G(p) E G ( P τ ∧ H ) = G ( p ) , while the concavity of G G G makes ( G ( P s ) ) s ∈ [ t , T ] (G(P_s))_{s\in[t,T]} ( G ( P s ) ) s ∈ [ t , T ] a super-martingale. Hence, since { H ≤ τ } ∈ F τ μ ∩ F H ∧ T μ \{H \le\tau\} \in\mathcal{F}^{\mu}_{\tau} \cap\mathcal{F}^{\mu}_{H \wedge T} { H ≤ τ } ∈ F τ μ ∩ F H ∧ T μ and P H ∧ T = m P_{H \wedge T}=m P H ∧ T = m on this event,
G ( p ) − E G ( P τ ) = E [ 1 { H ≤ τ } ( G ( m ) − G ( P τ ) ) ] ≤ E [ 1 { H ≤ τ } ( G ( m ) − G ( P T ) ) ] ≤ a ‾ c b 4 T − t P ( H ≤ τ ) , G(p)-\mathbb{E}G(P_{\tau})=\mathbb{E}\left[1_{\{H\le\tau\}}\left(G(m)-G(P_{\tau})\right)\right]
\le\mathbb{E}\left[1_{\{H\le\tau\}}\left(G(m)-G(P_T)\right)\right]
\le\frac{\overline{a}c_b}{4}\sqrt{T-t}\,\mathbb{P}(H\le\tau), G ( p ) − E G ( P τ ) = E [ 1 { H ≤ τ } ( G ( m ) − G ( P τ ) ) ] ≤ E [ 1 { H ≤ τ } ( G ( m ) − G ( P T ) ) ] ≤ 4 a c b T − t P ( H ≤ τ ) ,
with the last inequality following by conditioning at H ∧ T H\wedge T H ∧ T , using the Lipschitz bound for G G G , the conditional Itô isometry, and the Cauchy–Schwarz inequality.
Combining these estimates with f μ ≥ c f f^\mu\ge c_f f μ ≥ c f yields
E [ ∫ t τ f μ ( s , P s ) d s + G ( P τ ) ] − G ( p ) ≥ ( c f − a ‾ c b T − t 64 ∣ p − m ∣ 2 ) E ( τ − t ) . \mathbb{E}\left[\int_t^\tau f^\mu(s,P_s)\,\mathrm{d}s+G(P_\tau)\right]-G(p)
\ge\left(c_f-\frac{\overline{a}c_b\sqrt{T-t}}{64|p-m|^2}\right)\mathbb{E}(\tau-t). E [ ∫ t τ f μ ( s , P s ) d s + G ( P τ ) ] − G ( p ) ≥ ( c f − 64∣ p − m ∣ 2 a c b T − t ) E ( τ − t ) .
For every 0 < ε < min { m , 1 − m } 0<\varepsilon<\min\{m,1-m\} 0 < ε < min { m , 1 − m } , the bracket is nonnegative, uniformly in μ \mu μ , whenever ∣ p − m ∣ ≥ ε |p-m|\ge\varepsilon ∣ p − m ∣ ≥ ε and t t t is sufficiently close to T T T .
Since τ = t \tau=t τ = t is admissible, this proves V μ ( t , p ) = G ( p ) V^\mu(t,p)=G(p) V μ ( t , p ) = G ( p ) at all such points, and, thus Proposition 3.12 gives m − ε ≤ A − μ ( t ) < m < A + μ ( t ) ≤ m + ε m-\varepsilon\le A^\mu_-(t)<m<A^\mu_+(t)\le m+\varepsilon m − ε ≤ A − μ ( t ) < m < A + μ ( t ) ≤ m + ε , which proves the claim.
The existence of a mean field equilibrium
Here, we present the proof of the main result of the paper. For this we need to establish that, if μ n → μ \mu_n\to\mu μ n → μ in M \mathcal{M} M , then the best-response pair associated to μ n \mu_n μ n , as defined in (15), converges in probability to the best-response pair associated to μ \mu μ .
The first step is to prove that the signal and belief processes have this property.
Proposition 4.1. Assume (22) and let μ n , μ ∈ M \mu^n,\mu\in\mathcal{M} μ n , μ ∈ M be such that, as n → ∞ n\to\infty n → ∞ , μ n → μ \mu^n\to\mu μ n → μ . Consider the signal processes
d Y n ( t ) = b ( t , Θ ; F ( t ; ( μ n ) ≤ t ) ) d t + d W t Y n ( 0 ) = 0 , \mathrm{d}Y_n(t)=b(t,\Theta;F(t;(\mu^n)_{\le t}))\,\mathrm{d}t+\mathrm{d}W_t \qquad Y_n(0)=0, d Y n ( t ) = b ( t , Θ ; F ( t ; ( μ n ) ≤ t )) d t + d W t Y n ( 0 ) = 0 , and
d Y ( t ) = b ( t , Θ ; F ( t ; μ ≤ t ) ) d t + d W t Y ( 0 ) = 0 , \mathrm{d}Y(t)=b(t,\Theta;F(t;\mu_{\le t}))\,\mathrm{d}t+\mathrm{d}W_t \qquad Y(0)=0, d Y ( t ) = b ( t , Θ ; F ( t ; μ ≤ t )) d t + d W t Y ( 0 ) = 0 , and their respective belief processes P n : = P 0 , p 0 ; μ n P_n:=P^{0,p_0;\mu^n} P n := P 0 , p 0 ; μ n and P : = P 0 , p 0 ; μ P:=P^{0,p_0;\mu} P := P 0 , p 0 ; μ with common initial belief P 0 ∼ η P_0\sim\eta P 0 ∼ η , and set τ n : = τ ^ 0 , p 0 μ n \tau_n:=\hat{\tau}^{\mu^n}_{0,p_0} τ n := τ ^ 0 , p 0 μ n and τ : = τ ^ 0 , p 0 μ \tau:=\hat{\tau}^{\mu}_{0,p_0} τ := τ ^ 0 , p 0 μ . Then,
( τ n , P n ( τ n ) ) ⟶ ( τ , P ( τ ) ) in probability , (\tau_n,P_n(\tau_n))\longrightarrow(\tau,P(\tau))\quad\text{in probability}, ( τ n , P n ( τ n )) ⟶ ( τ , P ( τ )) in probability , and, consequently, Φ ( μ n ) → Φ ( μ ) \Phi(\mu^n)\to\Phi(\mu) Φ ( μ n ) → Φ ( μ ) in M \mathcal{M} M .
Proof. The belief processes P n P_n P n and P P P solve
d P n , t = ( Δ b ( t ; μ ≤ t n ) ) 2 P n , t ( 1 − P n , t ) ( Θ − P n , t ) d t + Δ b ( t ; μ ≤ t n ) P n , t ( 1 − P n , t ) d W t P n , 0 = P 0 , \mathrm{d}P_{n,t}=\left(\Delta b(t;\mu^n_{\le t})\right)^2P_{n,t}(1-P_{n,t})(\Theta-P_{n,t})\,\mathrm{d}t+\Delta b(t;\mu^n_{\le t})P_{n,t}(1-P_{n,t})\,\mathrm{d}W_t \qquad P_{n,0}=P_0, d P n , t = ( Δ b ( t ; μ ≤ t n ) ) 2 P n , t ( 1 − P n , t ) ( Θ − P n , t ) d t + Δ b ( t ; μ ≤ t n ) P n , t ( 1 − P n , t ) d W t P n , 0 = P 0 ,
and
d P t = ( Δ b ( t ; μ ≤ t ) ) 2 P t ( 1 − P t ) ( Θ − P t ) d t + Δ b ( t ; μ ≤ t ) P t ( 1 − P t ) d W t P 0 = P 0 . \mathrm{d}P_t=\left(\Delta b(t;\mu_{\le t})\right)^2P_t(1-P_t)(\Theta-P_t)\,\mathrm{d}t+\Delta b(t;\mu_{\le t})P_t(1-P_t)\,\mathrm{d}W_t \qquad P_0=P_0. d P t = ( Δ b ( t ; μ ≤ t ) ) 2 P t ( 1 − P t ) ( Θ − P t ) d t + Δ b ( t ; μ ≤ t ) P t ( 1 − P t ) d W t P 0 = P 0 .
Combining Lemma 3.1, the Burkholder-Davis-Grundy and Grönwall’s inequalities and Proposition 3.7, we find that, as n → ∞ n \to\infty n → ∞ ,
E [ sup t ≤ T ∣ P n ( t ) − P ( t ) ∣ 2 ] ⟶ 0 and ∥ V ^ μ n − V ^ μ ∥ ∞ ⟶ 0. \mathbb{E}\left[\sup_{t \leq T} \lvert P_n(t)-P(t)\rvert^2\right] \longrightarrow0 \quad\text{and} \quad\lVert\widehat{V}^{\mu^n}-\widehat{V}^{\mu}\rVert_{\infty} \longrightarrow0. E [ t ≤ T sup ∣ P n ( t ) − P ( t ) ∣ 2 ] ⟶ 0 and ∥ V μ n − V μ ∥ ∞ ⟶ 0.
Next, for t ∈ [ 0 , T ] t \in[0,T] t ∈ [ 0 , T ] , we write X n ( t ) : = − V ^ μ n ( t , P n , t ) X_n(t):=-\widehat{V}^{\mu^n}(t,P_{n,t}) X n ( t ) := − V μ n ( t , P n , t ) and X ( t ) : = − V ^ μ ( t , P t ) X(t):=-\widehat{V}^{\mu}(t,P_t) X ( t ) := − V μ ( t , P t ) , and denote by ω V ^ μ \omega_{\widehat{V}^{\mu}} ω V μ the modulus of continuity of V ^ μ \widehat{V}^{\mu} V μ on [ 0 , T ] × [ 0 , 1 ] [0,T]\times[0,1] [ 0 , T ] × [ 0 , 1 ] , which exists since V ^ μ \widehat{V}^{\mu} V μ is uniformly continuous on this compact set.
Then,
sup s ∈ [ 0 , T ] ∣ X n ( s ) − X ( s ) ∣ ≤ ∥ V ^ μ n − V ^ μ ∥ ∞ + ω V ^ μ ( ∥ P n − P ∥ ∞ ) → n → ∞ 0 in probability . (52) \sup_{s\in[0,T]}\lvert X_n(s)-X(s)\rvert\leq\lVert\widehat{V}^{\mu^n}-\widehat{V}^{\mu}\rVert_{\infty}+\omega_{\widehat{V}^{\mu}}\left(\lVert P_n-P\rVert_{\infty}\right) \xrightarrow[n\to\infty]{} 0 \quad\text{in probability}. \tag*{(52)} s ∈ [ 0 , T ] sup ∣ X n ( s ) − X ( s )∣ ≤ ∥ V μ n − V μ ∥ ∞ + ω V μ ( ∥ P n − P ∥ ∞ ) n → ∞ 0 in probability . (52)
Using the local uniform convergence on compact subsets of R ⊂ ( 0 , T ) \mathcal{R}\subset(0,T) R ⊂ ( 0 , T ) of the A ± μ n A_{\pm}^{\mu^n} A ± μ n ’s to A ± A_{\pm} A ± , the preceding L 2 L^2 L 2 convergence of P n P_n P n to P P P and (52) , we deduce that every subsequence admits a further subsequence, which, for notational simplicity, we denote the same as the sequence, and an event Ω 0 \Omega_0 Ω 0 with P ( Ω 0 ) = 1 \mathbb{P}(\Omega_0)=1 P ( Ω 0 ) = 1 such that, for every ω ∈ Ω 0 \omega\in\Omega_0 ω ∈ Ω 0 ,
sup s ∈ [ 0 , T ] ∣ P n , s ( ω ) − P s ( ω ) ∣ + sup s ∈ [ 0 , T ] ∣ X n ( s , ω ) − X ( s , ω ) ∣ → n → ∞ 0. (53) \sup_{s\in[0,T]}\lvert P_{n,s}(\omega)-P_s(\omega)\rvert+\sup_{s\in[0,T]}\lvert X_n(s,\omega)-X(s,\omega)\rvert\xrightarrow[n\to\infty]{}0. \tag*{(53)} s ∈ [ 0 , T ] sup ∣ P n , s ( ω ) − P s ( ω )∣ + s ∈ [ 0 , T ] sup ∣ X n ( s , ω ) − X ( s , ω )∣ n → ∞ 0. (53)
Let E 1 E_1 E 1 be a probability-one event on which P P P and every P n P_n P n have continuous paths; such an event exists because each process is continuous almost surely and the sequence is countable. Then, we replace Ω 0 \Omega_0 Ω 0 by Ω 0 ∩ E 1 \Omega_0\cap E_1 Ω 0 ∩ E 1 , which still has probability one, and we have that the convergence in (53) and path continuity hold on Ω 0 ∩ E 1 \Omega_0\cap E_1 Ω 0 ∩ E 1 .
Fix ω ∈ Ω 0 \omega\in\Omega_0 ω ∈ Ω 0 and δ > 0 \delta>0 δ > 0 . If τ ( ω ) > δ \tau(\omega)>\delta τ ( ω ) > δ , then
c δ ( ω ) : = min 0 ≤ s ≤ τ ( ω ) − δ X ( s , ω ) > 0. c_{\delta}(\omega):=\min_{0\leq s\leq\tau(\omega)-\delta}X(s,\omega)>0. c δ ( ω ) := 0 ≤ s ≤ τ ( ω ) − δ min X ( s , ω ) > 0.
Indeed, otherwise we would have X ( s , ω ) = 0 X(s,\omega)=0 X ( s , ω ) = 0 for some s ≤ τ ( ω ) − δ < τ ( ω ) s\leq\tau(\omega)-\delta<\tau(\omega) s ≤ τ ( ω ) − δ < τ ( ω ) , which contradicts the definition of τ ( ω ) \tau(\omega) τ ( ω ) .
Hence, for all sufficiently large n n n , X n ( s , ω ) ≥ c δ ( ω ) 2 X_n(s,\omega)\geq\frac{c_{\delta}(\omega)}{2} X n ( s , ω ) ≥ 2 c δ ( ω ) for 0 ≤ s ≤ τ ( ω ) − δ 0\leq s\leq\tau(\omega)-\delta 0 ≤ s ≤ τ ( ω ) − δ , and, thus, τ n ( ω ) ≥ τ ( ω ) − δ \tau_n(\omega)\geq\tau(\omega)-\delta τ n ( ω ) ≥ τ ( ω ) − δ , while, if τ ( ω ) ≤ δ \tau(\omega)\leq\delta τ ( ω ) ≤ δ , we obviously have τ n ( ω ) ≥ 0 ≥ τ ( ω ) − δ \tau_n(\omega)\geq0\geq\tau(\omega)-\delta τ n ( ω ) ≥ 0 ≥ τ ( ω ) − δ .
Letting δ → 0 \delta\to0 δ → 0 gives
lim inf n → ∞ τ n ( ω ) ≥ τ ( ω ) . (54) \liminf_{n\to\infty}\tau_n(\omega)\geq\tau(\omega). \tag*{(54)} n → ∞ lim inf τ n ( ω ) ≥ τ ( ω ) . (54)
Since ω ∈ Ω 0 \omega\in\Omega_0 ω ∈ Ω 0 was arbitrary and P ( Ω 0 ) = 1 \mathbb{P}(\Omega_0)=1 P ( Ω 0 ) = 1 , the inequality holds almost surely along the extracted subsequence.
We now prove that
lim sup n → ∞ τ n ≤ τ a.s. on { 0 ≤ τ ≤ T } . \limsup_{n\to\infty}\tau_n\leq\tau\quad\text{a.s. on } \{0\leq\tau\leq T\}. n → ∞ lim sup τ n ≤ τ a.s. on { 0 ≤ τ ≤ T } .
We begin with the proof of the claim in { 0 < τ < T } \{0<\tau<T\} { 0 < τ < T } .
Fix δ > 0 \delta>0 δ > 0 and recall that, in view of Proposition 3.17 , if τ ∈ ( 0 , T ) \tau\in(0,T) τ ∈ ( 0 , T ) is the first contact point of the process P P P , and depending on whether the contact is on A − μ A_-^{\mu} A − μ or A + μ A_+^{\mu} A + μ , the corresponding free boundary is C 0 , 1 / 2 C^{0,1/2} C 0 , 1/2 on ( τ − r ‾ , τ + r ‾ ) ∩ ( 0 , T ) (\tau-\overline r,\tau+\overline r)\cap(0,T) ( τ − r , τ + r ) ∩ ( 0 , T ) with constant C ‾ \overline C C , where C ‾ , r ‾ > 0 \overline C,\overline r>0 C , r > 0 may depend on μ \mu μ and the realized contact point.
Using the uniform lower bound of the a μ a^\mu a μ ’s and the law of iterated logarithm for Brownian motion (see, for example, [23 Theorem 2.9.23] ) we deduce that
lim inf h ↓ 0 P τ + h − P τ h = − ∞ and lim sup h ↓ 0 P τ + h − P τ h = ∞ . (55) \liminf_{h\downarrow0}\frac{P_{\tau+h}-P_\tau}{\sqrt{h}}=-\infty
\quad\text{and}\quad
\limsup_{h\downarrow0}\frac{P_{\tau+h}-P_\tau}{\sqrt{h}}=\infty. \tag*{(55)} h ↓ 0 lim inf h P τ + h − P τ = − ∞ and h ↓ 0 lim sup h P τ + h − P τ = ∞. (55)
Next, we show that (55) holds almost surely for a fixed μ \mu μ and τ \tau τ . Indeed, we apply the Dambis–Dubins–Schwarz theorem and the law of the iterated logarithm to M h : = P ( τ + h ) ∧ T − P τ M_h:=P_{(\tau+h)\wedge T}-P_\tau M h := P ( τ + h ) ∧ T − P τ , which is a martingale with respect to ( F ( τ + h ) ∧ T μ ) h ≥ 0 (\mathcal F^\mu_{(\tau+h)\wedge T})_{h\geq0} ( F ( τ + h ) ∧ T μ ) h ≥ 0 with quadratic variation ⟨ M ⟩ \langle M\rangle ⟨ M ⟩ satisfying, on { 0 < τ < T } \{0<\tau<T\} { 0 < τ < T } ,
⟨ M ⟩ h / h ⟶ ( Δ b μ ( τ ) ) 2 P τ 2 ( 1 − P τ ) 2 > 0. \langle M\rangle_h/h\longrightarrow(\Delta b^\mu(\tau))^2P_\tau^2(1-P_\tau)^2>0. ⟨ M ⟩ h / h ⟶ ( Δ b μ ( τ ) ) 2 P τ 2 ( 1 − P τ ) 2 > 0.
In view of Proposition 3.17 , there is a probability-one event E 2 E_2 E 2 on which, whenever 0 < τ < T 0<\tau<T 0 < τ < T , first contact occurs at one of the two boundaries and the contacted boundary is locally C 1 / 2 C^{1/2} C 1/2 . The above argument gives a probability-one event E 3 E_3 E 3 on which both limits in (55) hold whenever 0 < τ < T 0<\tau<T 0 < τ < T . Replace Ω 0 \Omega_0 Ω 0 by Ω 0 ∩ E 2 ∩ E 3 \Omega_0\cap E_2\cap E_3 Ω 0 ∩ E 2 ∩ E 3 , which still has probability one. The convergence in (53) and the lim inf \liminf lim inf bound in (54) remain valid on this smaller event.
Fix ω ∈ Ω 0 ∩ { 0 < τ < T } \omega\in\Omega_0\cap\{0<\tau<T\} ω ∈ Ω 0 ∩ { 0 < τ < T } and δ > 0 \delta>0 δ > 0 . For notational simplicity in what follows we suppress ω \omega ω .
In what follows, we examine separately the cases where the contact point is either on the lower (left) free boundary A − μ A^\mu_- A − μ or upper (right) free boundary A + μ A^\mu_+ A + μ .
If the former holds, using (55) , we can choose h > 0 h>0 h > 0 small enough (and smaller than δ \delta δ and r ‾ \overline r r in Proposition 3.17 ) to obtain
P τ + h − P τ < − C ‾ h . P_{\tau+h}-P_\tau<-\overline C\sqrt{h}. P τ + h − P τ < − C h .
It follows from the local C 1 / 2 C^{1/2} C 1/2 -continuity of the interfaces with constant C ‾ \overline C C that
P τ + h < A − μ ( τ ) − C ‾ h ≤ A − μ ( τ + h ) . P_{\tau+h}<A^\mu_-(\tau)-\overline C\sqrt{h}\leq A^\mu_-(\tau+h). P τ + h < A − μ ( τ ) − C h ≤ A − μ ( τ + h ) .
If the contact is on A + μ A^\mu_+ A + μ , then, in view of (55) , we choose h > 0 h>0 h > 0 small enough (and smaller than δ \delta δ and r ‾ \overline r r in Proposition 3.17 ) so that
P τ + h − P τ > C ‾ h . P_{\tau+h}-P_\tau>\overline C\sqrt{h}. P τ + h − P τ > C h .
Then, using the local C 1 / 2 C^{1/2} C 1/2 -continuity of the interfaces with constant C ‾ \overline C C , we get
P τ + h > A + μ ( τ ) + C ‾ h ≥ A + μ ( τ + h ) . P_{\tau+h}>A^\mu_+(\tau)+\overline C\sqrt{h}\geq A^\mu_+(\tau+h). P τ + h > A + μ ( τ ) + C h ≥ A + μ ( τ + h ) .
In summary, there exists a nonempty compact set J ⊂ ( τ , ( τ + δ ) ∧ T ) ∩ R J\subset(\tau,(\tau+\delta)\wedge T)\cap\mathbb R J ⊂ ( τ , ( τ + δ ) ∧ T ) ∩ R on which
either P s < A − μ ( s ) or P s > A + μ ( s ) , \text{either }P_s<A^\mu_-(s)\quad\text{or}\quad P_s>A^\mu_+(s), either P s < A − μ ( s ) or P s > A + μ ( s ) ,
while, as shown in the previous section, as n → ∞ n\to\infty n → ∞ and uniformly on J J J ,
A − μ n ⟶ A − μ and A + μ n ⟶ A + μ . A^{\mu_n}_-\longrightarrow A^\mu_-
\quad\text{and}\quad
A^{\mu_n}_+\longrightarrow A^\mu_+. A − μ n ⟶ A − μ and A + μ n ⟶ A + μ .
It then follows that, for all large n n n ,
either P n ( s ) < A − μ n ( s ) or P n ( s ) > A + μ n ( s ) , \text{either }P_n(s)<A^{\mu_n}_-(s)\quad\text{or}\quad P_n(s)>A^{\mu_n}_+(s), either P n ( s ) < A − μ n ( s ) or P n ( s ) > A + μ n ( s ) ,
and, thus, τ n ≤ s < τ + δ \tau_n \le s < \tau+ \delta τ n ≤ s < τ + δ , and, in turn, lim sup n → ∞ τ n ≤ τ \limsup_{n\to\infty} \tau_n \le\tau lim sup n → ∞ τ n ≤ τ on this event.
Since on { τ = T } \{\tau= T\} { τ = T } , it holds that τ n ≤ T = τ \tau_n \le T = \tau τ n ≤ T = τ , it only remains to show that on { τ = 0 } \{\tau= 0\} { τ = 0 } we also have lim sup n → ∞ τ n ≤ τ \limsup_{n\to\infty} \tau_n \le\tau lim sup n → ∞ τ n ≤ τ .
Let R \mathcal{R} R be the full measure set from Theorem 3.15 and choose s j ∈ R s_j \in\mathcal{R} s j ∈ R with s j ↓ 0 s_j \downarrow0 s j ↓ 0 . On { τ = 0 } \{\tau= 0\} { τ = 0 } , Proposition 3.12 gives
p 0 ∈ [ 0 , A − μ ( 0 ) ] ∪ [ A + μ ( 0 ) , 1 ] . p_0 \in[0,A^\mu_-(0)] \cup[A^\mu_+(0),1]. p 0 ∈ [ 0 , A − μ ( 0 )] ∪ [ A + μ ( 0 ) , 1 ] .
Since η \eta η has a density, then either p 0 < A − μ ( 0 ) p_0 < A^\mu_-(0) p 0 < A − μ ( 0 ) or p 0 > A + μ ( 0 ) p_0 > A^\mu_+(0) p 0 > A + μ ( 0 ) holds almost surely.
We only consider the case p 0 < A − μ ( 0 ) p_0 < A^\mu_-(0) p 0 < A − μ ( 0 ) , as the other one follows similarly.
In view of the continuity of the belief process and Proposition 3.19 , there exists r > 0 r > 0 r > 0 such that P s < A − μ ( s ) P_s < A^\mu_-(s) P s < A − μ ( s ) for 0 ≤ s < r 0 \le s < r 0 ≤ s < r , and, for every sufficiently large j j j , s j < r s_j < r s j < r , and, furthermore,
P n , s j → P s j and A − μ n ( s j ) → A − μ ( s j ) . P_{n,s_j} \to P_{s_j} \quad\text{and} \quad A^{\mu_n}_-(s_j) \to A^\mu_-(s_j). P n , s j → P s j and A − μ n ( s j ) → A − μ ( s j ) .
Hence, P n , s j < A − μ n ( s j ) P_{n,s_j} < A^{\mu_n}_-(s_j) P n , s j < A − μ n ( s j ) for all large n n n , and, thus, τ n ≤ s j \tau_n \le s_j τ n ≤ s j .
It then follows that lim sup n τ n ≤ s j \limsup_{n} \tau_n \le s_j lim sup n τ n ≤ s j for every sufficiently large j j j , and letting j → ∞ j \to\infty j → ∞ we deduce
lim sup n → ∞ τ n ≤ 0 = τ . \limsup_{n\to\infty} \tau_n \le0 = \tau. n → ∞ lim sup τ n ≤ 0 = τ .
Next, we prove the convergence of the decision rules D ( P n , τ n ) D(P_{n,\tau_n}) D ( P n , τ n ) .
If τ < T \tau< T τ < T , Proposition 3.12 gives P τ ∈ [ 0 , A − μ ( τ ) ] ∪ [ A + μ ( τ ) , 1 ] P_\tau\in[0,A^\mu_-(\tau)] \cup[A^\mu_+(\tau),1] P τ ∈ [ 0 , A − μ ( τ )] ∪ [ A + μ ( τ ) , 1 ] , which does not contain m m m . If τ = T \tau= T τ = T , then P T P_T P T has no atom at m m m , because the logit diffusion has a transition density on ( 0 , 1 ) (0,1) ( 0 , 1 ) . If τ = 0 \tau= 0 τ = 0 , then, as before, it suffices to consider the cases that p p p is strictly either below A − μ ( 0 ) A^\mu_-(0) A − μ ( 0 ) or above A + μ ( 0 ) A^\mu_+(0) A + μ ( 0 ) and, hence, away from m m m .
Thus, the decision rule is continuous at P τ P_\tau P τ almost surely in all cases, and, as n → ∞ n \to\infty n → ∞ , D ( P n , τ n ) ⟶ D ( P τ ) D(P_{n,\tau_n}) \longrightarrow D(P_\tau) D ( P n , τ n ) ⟶ D ( P τ ) in probability.
Finally, for h ∈ C ( [ 0 , T ] ) h \in C([0,T]) h ∈ C ([ 0 , T ]) and i ∈ { 0 , 1 } i \in\{0,1\} i ∈ { 0 , 1 } , we have
∫ [ 0 , T ] h ( t ) ν i μ n ( d t ) = E [ h ( τ 0 , P 0 μ n ) 1 { D ( P n , τ n ) = i } ] . \int_{[0,T]} h(t)\,\nu_i^{\mu_n}(\mathrm{d}t)=\mathbb{E}\left[h(\tau^{\mu_n}_{0,P_0})1_{\{D(P_{n,\tau_n})=i\}}\right]. ∫ [ 0 , T ] h ( t ) ν i μ n ( d t ) = E [ h ( τ 0 , P 0 μ n ) 1 { D ( P n , τ n ) = i } ] .
The right-hand side converges to the analogous expression for μ \mu μ , which proves Φ ( μ n ) ⇀ Φ ( μ ) \Phi(\mu^n) \rightharpoonup\Phi(\mu) Φ ( μ n ) ⇀ Φ ( μ ) .
We conclude with the proof of the existence of the mean field equilibria.
Theorem 4.2. Assume (22). Then, there exists μ ∗ ∈ M \mu^* \in\mathcal{M} μ ∗ ∈ M such that Φ ( μ ∗ ) = μ ∗ \Phi(\mu^*) = \mu^* Φ ( μ ∗ ) = μ ∗ .
Proof. The set M \mathcal{M} M is nonempty, convex and compact with respect to the weak topology, while Proposition 4.1 shows that Φ : M → M \Phi:\mathcal{M}\to\mathcal{M} Φ : M → M is continuous. Then, it follows from the Schauder–Tychonoff fixed-point theorem that there exists μ ∗ ∈ M \mu^* \in\mathcal{M} μ ∗ ∈ M with Φ ( μ ∗ ) = μ ∗ \Phi(\mu^*) = \mu^* Φ ( μ ∗ ) = μ ∗ .
The finite game
We consider the finite analogue of the information acquisition mean field game and examine the proximity of its solution to the ones in the continuum, constructed in the previous section. As mentioned in the introduction, we identify each decision maker (DM) with the pair of his prior belief and the related posterior belief process, generated by the signal whose dynamics depend on the stopping times, parameter assessment and costs of the rest of the populations. In this sense, there are no players as such in the standard way players are considered in mean field games. For this, we choose to use the terminology “decision maker” rather than a player, and we map any DM to the corresponding probabilities (prior and posterior).
Fix N ≥ 2 N \ge2 N ≥ 2 . DM j j j has initial belief p j ∼ η p^{j} \sim\eta p j ∼ η and a private innovation Brownian motion Z j Z^{j} Z j independent of p j p^{j} p j , with the pairs ( p j , Z j ) (p^{j}, Z^{j}) ( p j , Z j ) independent across DMs.
Let Ω 0 = C 0 ( [ 0 , T ] ; R ) \Omega_{0} = C_{0}([0,T];\mathbb{R}) Ω 0 = C 0 ([ 0 , T ] ; R ) with Wiener measure W \mathbb{W} W , and set
P N : = ( η ⊗ W ) ⊗ N on ( [ 0 , 1 ] × Ω 0 ) N . \mathbb{P}^{N} := (\eta\otimes\mathbb{W})^{\otimes N} \quad\text{on} \quad([0,1] \times\Omega_{0})^{N}. P N := ( η ⊗ W ) ⊗ N on ([ 0 , 1 ] × Ω 0 ) N .
For a fixed player i i i and initial belief p p p , let P i , p N \mathbb{P}_{i,p}^{N} P i , p N be the same product law with p i = p p^{i}=p p i = p , and write E i , p N \mathbb{E}_{i,p}^{N} E i , p N for expectation under this law.
Let μ ∗ \mu^{*} μ ∗ be a mean field equilibrium constructed in Theorem 4.2 , and, for each j = 1 , … , N j=1,\ldots,N j = 1 , … , N , consider best response strategy pair
( τ ˉ N , j , d ˉ N , j ) : = ( τ ^ 0 , p j μ ∗ , d ^ 0 , p j μ ∗ ) , (56) (\bar{\tau}^{N,j},\bar{d}^{N,j}) := (\hat{\tau}_{0,p^{j}}^{\mu^{*}},\hat{d}_{0,p^{j}}^{\mu^{*}}), \tag*{(56)} ( τ ˉ N , j , d ˉ N , j ) := ( τ ^ 0 , p j μ ∗ , d ^ 0 , p j μ ∗ ) , (56)
with the belief process driven by Z j Z^{j} Z j .
Since Φ ( μ ∗ ) = μ ∗ \Phi(\mu^{*})=\mu^{*} Φ ( μ ∗ ) = μ ∗ , for every Borel A ⊂ [ 0 , T ] A \subset[0,T] A ⊂ [ 0 , T ] and k = 0 , 1 k=0,1 k = 0 , 1 , we deduce that, for each j = 1 , … , N j=1,\ldots,N j = 1 , … , N ,
P N ( τ ˉ N , j ∈ A , d ˉ N , j = k ) = μ ∗ , k ( A ) . \mathbb{P}^{N}(\bar{\tau}^{N,j} \in A,\bar{d}^{N,j}=k)=\mu^{*,k}(A). P N ( τ ˉ N , j ∈ A , d ˉ N , j = k ) = μ ∗ , k ( A ) .
Therefore, all pairs ( τ ˉ N , j , d ˉ N , j ) (\bar{\tau}^{N,j},\bar{d}^{N,j}) ( τ ˉ N , j , d ˉ N , j ) are independent and all have the equilibrium distribution μ ∗ \mu^{*} μ ∗ .
Next, we fix i i i . For j = 1 , … , i − 1 , i + 1 , … , N j=1,\ldots,i-1,i+1,\ldots,N j = 1 , … , i − 1 , i + 1 , … , N , we consider the pairs given by (56) and, for k = 0 , 1 k=0,1 k = 0 , 1 , we introduce the measures
μ ^ − i , N , k : = 1 N − 1 ∑ j ≠ i 1 { d ˉ N , j = k } δ τ ˉ N , j and μ ^ − i , N : = ( μ ^ − i , N , 0 , μ ^ − i , N , 1 ) . \hat{\mu}^{-i,N,k} := \frac{1}{N-1}\sum_{j\ne i}\mathbf{1}_{\{\bar{d}^{N,j}=k\}}\delta_{\bar{\tau}^{N,j}}
\quad\text{and}\quad
\hat{\mu}^{-i,N}:=(\hat{\mu}^{-i,N,0},\hat{\mu}^{-i,N,1}). μ ^ − i , N , k := N − 1 1 j = i ∑ 1 { d ˉ N , j = k } δ τ ˉ N , j and μ ^ − i , N := ( μ ^ − i , N , 0 , μ ^ − i , N , 1 ) .
Let ( F t N , i , p ) (\mathcal{F}_{t}^{N,i,p}) ( F t N , i , p ) be the right continuous completion under P i , p N \mathbb{P}_{i,p}^{N} P i , p N of the σ \sigma σ -algebra
σ ( p i , Z s i , μ ^ − i , N , k ( [ 0 , s ] ) : 0 ≤ s ≤ t , k = 0 , 1 ) . \sigma\left(p^{i},Z_{s}^{i},\hat{\mu}^{-i,N,k}([0,s]):0\le s\le t,\ k=0,1\right). σ ( p i , Z s i , μ ^ − i , N , k ([ 0 , s ]) : 0 ≤ s ≤ t , k = 0 , 1 ) .
It then follows from the discussion in § 2 that the map t ↦ F ( t ; μ ^ ≤ t − i , N ) t \mapsto F(t;\hat{\mu}_{\le t}^{-i,N}) t ↦ F ( t ; μ ^ ≤ t − i , N ) is continuous and adapted, and, hence, predictable.
Next, we introduce the processes
d P t i , N , p = Δ b ( t ; μ ^ ≤ t − i , N ) P t i , N , p ( 1 − P t i , N , p ) d Z t i , P 0 i , N , p = p , (57) \mathrm{d}P_{t}^{i,N,p}=\Delta b(t;\hat{\mu}_{\le t}^{-i,N})P_{t}^{i,N,p}(1-P_{t}^{i,N,p})\,\mathrm{d}Z_{t}^{i},
\qquad P_{0}^{i,N,p}=p, \tag*{(57)} d P t i , N , p = Δ b ( t ; μ ^ ≤ t − i , N ) P t i , N , p ( 1 − P t i , N , p ) d Z t i , P 0 i , N , p = p , (57)
d P t i , ∗ , p = Δ b ( t ; μ ≤ t ∗ ) P t i , ∗ , p ( 1 − P t i , ∗ , p ) d Z t i , P 0 i , ∗ , p = p , (58) \mathrm{d}P_{t}^{i,*,p}=\Delta b(t;\mu_{\le t}^{*})P_{t}^{i,*,p}(1-P_{t}^{i,*,p})\,\mathrm{d}Z_{t}^{i},
\qquad P_{0}^{i,*,p}=p, \tag*{(58)} d P t i , ∗ , p = Δ b ( t ; μ ≤ t ∗ ) P t i , ∗ , p ( 1 − P t i , ∗ , p ) d Z t i , P 0 i , ∗ , p = p , (58)
which represent the posterior empirical environment and its mean field comparison process.
Since the coefficients above are predictable and uniformly Lipschitz in the state variable, (57) and (58) have unique strong F N , i , p \mathcal{F}^{N,i,p} F N , i , p -adapted solutions.
Let A i N , p \mathcal{A}_{i}^{N,p} A i N , p be the pairs ( τ , d ) (\tau,d) ( τ , d ) , with τ \tau τ and d d d being respectively an [ 0 , T ] [0,T] [ 0 , T ] -valued F N , i , p \mathcal{F}^{N,i,p} F N , i , p -stopping time and F τ N , i , p \mathcal{F}_{\tau}^{N,i,p} F τ N , i , p -measurable.
Then, in complete analogy to (6) , we consider the cost
J i N ( p ; τ , d ) : = E i , p N [ ∫ 0 τ f ( s , P s i , N , p ; μ ^ ≤ s − i , N ) d s + G ‾ ( P τ i , N , p , d ) ] . J_i^N(p;\tau,d) := \mathbb{E}_{i,p}^N\left[\int_0^\tau f(s,P_s^{i,N,p};\widehat{\mu}_{\leq s}^{-i,N})\,\mathrm{d}s+\overline{G}(P_\tau^{i,N,p},d)\right]. J i N ( p ; τ , d ) := E i , p N [ ∫ 0 τ f ( s , P s i , N , p ; μ ≤ s − i , N ) d s + G ( P τ i , N , p , d ) ] .
Finally, under P i , p N \mathbb{P}_{i,p}^N P i , p N , we define
τ ‾ N , i = inf { t ∈ [ 0 , T ] : V μ ∗ ( t , P t i , ∗ , p ) = G ( P t i , ∗ , p ) } and d ‾ N , i = D ( P τ ‾ N , i i , ∗ , p ) , \overline{\tau}^{N,i}=\inf\{t\in[0,T]:V\mu^*(t,P_t^{i,*,p})=G(P_t^{i,*,p})\}\quad\text{and}\quad\overline{d}^{N,i}=D(P_{\overline{\tau}^{N,i}}^{i,*,p}), τ N , i = inf { t ∈ [ 0 , T ] : V μ ∗ ( t , P t i , ∗ , p ) = G ( P t i , ∗ , p )} and d N , i = D ( P τ N , i i , ∗ , p ) ,
and easily deduce that ( τ ‾ N , i , d ‾ N , i ) ∈ A i N , p (\overline{\tau}^{N,i},\overline{d}^{N,i})\in\mathcal{A}_{i}^{N,p} ( τ N , i , d N , i ) ∈ A i N , p .
The result about the finite size approximation is stated and proved next.
Theorem 5.1. Let μ ∗ \mu^* μ ∗ be the mean field equilibrium constructed in Theorem 4.2 . Then, there exists a constant C > 0 C>0 C > 0 , which is independent of N , i N,i N , i , and p p p , such that, for every p ∈ [ 0 , 1 ] p\in[0,1] p ∈ [ 0 , 1 ] , N ≥ 2 N\geq2 N ≥ 2 , i = 1 , … , N i=1,\ldots,N i = 1 , … , N , and ( τ , d ) ∈ A i N , p (\tau,d)\in\mathcal{A}_{i}^{N,p} ( τ , d ) ∈ A i N , p ,
J i N ( p ; τ ‾ N , i , d ‾ N , i ) ≤ J i N ( p ; τ , d ) + C N − 1 . J_i^N(p;\overline{\tau}^{N,i},\overline{d}^{N,i})\leq J_i^N(p;\tau,d)+\frac{C}{\sqrt{N-1}}. J i N ( p ; τ N , i , d N , i ) ≤ J i N ( p ; τ , d ) + N − 1 C . Consequently, the mean field solution is an ε N \varepsilon_N ε N -Nash equilibrium with ε N = C / N − 1 \varepsilon_N=C/\sqrt{N-1} ε N = C / N − 1 .
Proof. Throughout the proof, C C C denotes a generic constant independent of N , p N,p N , p and i = 1 , … , N i=1,\ldots,N i = 1 , … , N .
For fixed k = 0 , 1 k=0,1 k = 0 , 1 and j = 1 , … , N j=1,\ldots,N j = 1 , … , N such that j ≠ i j\ne i j = i , we define
T j k : = { τ ‾ N , j , if d ‾ N , j = k , T + 1 , if d ‾ N , j ≠ k . T_j^k:=\begin{cases}
\overline{\tau}^{N,j}, & \text{if }\overline{d}^{N,j}=k,\\
T+1, & \text{if }\overline{d}^{N,j}\ne k.
\end{cases} T j k := { τ N , j , T + 1 , if d N , j = k , if d N , j = k .
It follows that ( T j k ) j ≠ i (T_j^k)_{j\ne i} ( T j k ) j = i are i.i.d. and, for t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
μ ^ − i , N , t k = 1 N − 1 ∑ j ≠ i 1 { T j k ≤ t } and P ( T j k ≤ t ) = μ ∗ , k ( [ 0 , t ] ) . \widehat{\mu}_{-i,N,t}^k=\frac{1}{N-1}\sum_{j\ne i}\mathbf{1}_{\{T_j^k\leq t\}}
\quad\text{and}\quad
\mathbb{P}(T_j^k\leq t)=\mu^{*,k}([0,t]). μ − i , N , t k = N − 1 1 j = i ∑ 1 { T j k ≤ t } and P ( T j k ≤ t ) = μ ∗ , k ([ 0 , t ]) .
Thus, t ↦ μ ^ − i , N , t k t\mapsto\widehat{\mu}_{-i,N,t}^k t ↦ μ − i , N , t k is the corresponding empirical distribution function, and the Dvoretzky–Kiefer–Wolfowitz inequality (see [26 ] ) yields
E i , p N [ sup t ≤ T ∣ μ ^ − i , N k ( [ 0 , t ] ) − μ ∗ , k ( [ 0 , t ] ) ∣ 2 ] ≤ 1 N − 1 . (59) \mathbb{E}_{i,p}^N\left[\sup_{t\leq T}\left|\widehat{\mu}_{-i,N}^k([0,t])-\mu^{*,k}([0,t])\right|^2\right]\leq\frac{1}{N-1}. \tag*{(59)} E i , p N [ t ≤ T sup μ − i , N k ([ 0 , t ]) − μ ∗ , k ([ 0 , t ]) 2 ] ≤ N − 1 1 . (59)
Moreover (16) and Fubini’s theorem imply that, for every finite measure ν \nu ν ,
F k ( t ; ν ≤ t ) = ∫ 0 t ν ( [ 0 , s ] ) ϕ ( t − s ) d s . F_k(t;\nu_{\leq t})=\int_0^t\nu([0,s])\phi(t-s)\,\mathrm{d}s. F k ( t ; ν ≤ t ) = ∫ 0 t ν ([ 0 , s ]) ϕ ( t − s ) d s .
Therefore,
sup t ≤ T ∣ F k ( t ; μ ^ ≤ t − i , N , k ) − F k ( t ; μ ≤ t ∗ , k ) ∣ ≤ sup t ≤ T ∣ μ ^ − i , N k ( [ 0 , t ] ) − μ ∗ , k ( [ 0 , t ] ) ∣ , \sup_{t\leq T}\left|F_k(t;\widehat{\mu}_{\leq t}^{-i,N,k})-F_k(t;\mu_{\leq t}^{*,k})\right|
\leq\sup_{t\leq T}\left|\widehat{\mu}_{-i,N}^k([0,t])-\mu^{*,k}([0,t])\right|, t ≤ T sup F k ( t ; μ ≤ t − i , N , k ) − F k ( t ; μ ≤ t ∗ , k ) ≤ t ≤ T sup μ − i , N k ([ 0 , t ]) − μ ∗ , k ([ 0 , t ]) ,
and, in turn, (60) implies
E i , p N [ sup t ≤ T ∣ F ( t ; μ ^ ≤ t − i , N ) − F ( t ; μ ≤ t ∗ ) ∣ 2 ] ≤ 4 N − 1 . (60) \mathbb{E}^{N}_{i,p}\left[\sup_{t\leq T}\left|F(t;\widehat{\mu}^{-i,N}_{\leq t})-F(t;\mu^{*}_{\leq t})\right|^{2}\right]\leq\frac{4}{N-1}. \tag*{(60)} E i , p N [ t ≤ T sup F ( t ; μ ≤ t − i , N ) − F ( t ; μ ≤ t ∗ ) 2 ] ≤ N − 1 4 . (60)
For subprobability measures ν , ν ′ \nu,\nu' ν , ν ′ on [ 0 , T ] [0,T] [ 0 , T ] , Kantorovich duality and integration by parts give
W 1 e x t ( ν , ν ′ ) ≤ ( 2 T + 1 ) sup t ∈ [ 0 , T ] ∣ ν ( [ 0 , t ] ) − ν ′ ( [ 0 , t ] ) ∣ . W^{\mathrm{ext}}_{1}(\nu,\nu')\leq(2T+1)\sup_{t\in[0,T]}\left|\nu([0,t])-\nu'([0,t])\right|. W 1 ext ( ν , ν ′ ) ≤ ( 2 T + 1 ) t ∈ [ 0 , T ] sup ∣ ν ([ 0 , t ]) − ν ′ ([ 0 , t ]) ∣ .
Thus, (60) and the Lipschitz assumption (18) imply
E i , p N [ sup t ≤ T ∣ Δ b ( t ; μ ^ ≤ t − i , N ) − Δ b ( t ; μ ≤ t ∗ ) ∣ 2 ] ≤ C E i , p N [ d M ( μ ^ − i , N , μ ∗ ) 2 ] ≤ C N − 1 . \mathbb{E}^{N}_{i,p}\left[\sup_{t\leq T}\left|\Delta b(t;\widehat{\mu}^{-i,N}_{\leq t})-\Delta b(t;\mu^{*}_{\leq t})\right|^{2}\right]
\leq C\mathbb{E}^{N}_{i,p}\left[d_{\mathcal{M}}(\widehat{\mu}^{-i,N},\mu^{*})^{2}\right]\leq\frac{C}{N-1}. E i , p N [ t ≤ T sup Δ b ( t ; μ ≤ t − i , N ) − Δ b ( t ; μ ≤ t ∗ ) 2 ] ≤ C E i , p N [ d M ( μ − i , N , μ ∗ ) 2 ] ≤ N − 1 C .
It follows from the Lipschitz continuity of the coefficients of (58) and (59), the Burkholder–Davis–Grundy inequality, and the above estimate, that
E i , p N [ sup r < t ∣ P r i , N , p − P r i , ∗ , p ∣ 2 ] ≤ C ∫ 0 t E i , p N [ sup u ≤ s ∣ P u i , N , p − P u i , ∗ , p ∣ 2 ] d s + C N − 1 . \mathbb{E}^{N}_{i,p}\left[\sup_{r<t}\left|P^{i,N,p}_{r}-P^{i,*,p}_{r}\right|^{2}\right]\leq C\int_{0}^{t}\mathbb{E}^{N}_{i,p}\left[\sup_{u\leq s}\left|P^{i,N,p}_{u}-P^{i,*,p}_{u}\right|^{2}\right]\mathrm{d}s+\frac{C}{N-1}. E i , p N [ r < t sup P r i , N , p − P r i , ∗ , p 2 ] ≤ C ∫ 0 t E i , p N [ u ≤ s sup P u i , N , p − P u i , ∗ , p 2 ] d s + N − 1 C .
In turn, using Grönwall’s inequality, we find that, uniformly in i i i and p p p ,
E i , p N [ sup t ≤ T ∣ P t i , N , p − P t i , ∗ , p ∣ 2 ] ≤ C N − 1 . (61) \mathbb{E}^{N}_{i,p}\left[\sup_{t\leq T}\left|P^{i,N,p}_{t}-P^{i,*,p}_{t}\right|^{2}\right]\leq\frac{C}{N-1}. \tag*{(61)} E i , p N [ t ≤ T sup P t i , N , p − P t i , ∗ , p 2 ] ≤ N − 1 C . (61)
For any ( τ , d ) ∈ A i N , p (\tau,d)\in\mathcal{A}^{N,p}_{i} ( τ , d ) ∈ A i N , p , its cost under the deterministic environment μ ∗ \mu^{*} μ ∗ is given by
J i ∗ ( p ; τ , d ) : = E i , p N [ ∫ 0 τ f ( s , P s i , ∗ , p ; μ ≤ s ∗ ) d s + G ‾ ( P τ i , ∗ , p , d ) ] . J^{*}_{i}(p;\tau,d):=\mathbb{E}^{N}_{i,p}\left[\int_{0}^{\tau}f(s,P^{i,*,p}_{s};\mu^{*}_{\leq s})\,\mathrm{d}s+\overline{G}(P^{i,*,p}_{\tau},d)\right]. J i ∗ ( p ; τ , d ) := E i , p N [ ∫ 0 τ f ( s , P s i , ∗ , p ; μ ≤ s ∗ ) d s + G ( P τ i , ∗ , p , d ) ] .
The Lipschitz bounds on f f f and G G G , together with (61), (62), and Cauchy–Schwarz inequality, give the uniform over ( τ , d ) ∈ A i N , p (\tau,d)\in\mathcal{A}^{N,p}_{i} ( τ , d ) ∈ A i N , p estimate
∣ J i N ( p ; τ , d ) − J i ∗ ( p ; τ , d ) ∣ ≤ C E i , p N [ sup t ≤ T ∣ P t i , N , p − P t i , ∗ , p ∣ + sup t ≤ T ∣ F ( t ; μ ^ ≤ t − i , N ) − F ( t ; μ ≤ t ∗ ) ∣ ] ≤ C N − 1 . (62) \left|J^{N}_{i}(p;\tau,d)-J^{*}_{i}(p;\tau,d)\right|\leq C\mathbb{E}^{N}_{i,p}\left[\sup_{t\leq T}\left|P^{i,N,p}_{t}-P^{i,*,p}_{t}\right|+\sup_{t\leq T}\left|F(t;\widehat{\mu}^{-i,N}_{\leq t})-F(t;\mu^{*}_{\leq t})\right|\right]
\leq\frac{C}{\sqrt{N-1}}. \tag*{(62)} J i N ( p ; τ , d ) − J i ∗ ( p ; τ , d ) ≤ C E i , p N [ t ≤ T sup P t i , N , p − P t i , ∗ , p + t ≤ T sup F ( t ; μ ≤ t − i , N ) − F ( t ; μ ≤ t ∗ ) ] ≤ N − 1 C . (62)
Finally, the information generated by the belief processes does not lower the value in the fixed environment μ ∗ \mu^{*} μ ∗ . Indeed, we fix their sample paths, and note that, for almost every realization ω − i \omega^{-i} ω − i if τ ( ⋅ , ω − i ) \tau(\cdot,\omega^{-i}) τ ( ⋅ , ω − i ) is a stopping time for the private filtration of P i P^{i} P i , then, for k = 0 , 1 k=0,1 k = 0 , 1 and t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] ,
{ d ( ⋅ , ω − i ) = k } ∩ { τ ( ⋅ , ω − i ) ≤ t } ∈ σ ( p i , Z s i : 0 ≤ s ≤ t ) . \{d(\cdot,\omega^{-i})=k\}\cap\{\tau(\cdot,\omega^{-i})\leq t\}\in\sigma(p^{i},Z^{i}_{s}:0\leq s\leq t). { d ( ⋅ , ω − i ) = k } ∩ { τ ( ⋅ , ω − i ) ≤ t } ∈ σ ( p i , Z s i : 0 ≤ s ≤ t ) .
Therefore, for almost every ω − i \omega^{-i} ω − i , fixing ω − i \omega^{-i} ω − i makes ( τ ( ⋅ , ω − i ) , d ( ⋅ , ω − i ) ) (\tau(\cdot,\omega^{-i}),d(\cdot,\omega^{-i})) ( τ ( ⋅ , ω − i ) , d ( ⋅ , ω − i )) admissible with respect to this filtration.
On the other hand, the canonical pair is optimal in the environment μ ∗ \mu^* μ ∗ , we deduce that
J i ∗ ( p ; τ ˉ N , i , d ˉ N , i ) = V μ ∗ ( 0 , p ) ≤ J i ∗ ( p ; τ ( ⋅ , ω − i ) , d ( ⋅ , ω − i ) ) . J_i^*(p;\bar{\tau}^{N,i},\bar{d}^{N,i})=V^{\mu^*}(0,p)\leq J_i^*(p;\tau(\cdot,\omega^{-i}),d(\cdot,\omega^{-i})). J i ∗ ( p ; τ ˉ N , i , d ˉ N , i ) = V μ ∗ ( 0 , p ) ≤ J i ∗ ( p ; τ ( ⋅ , ω − i ) , d ( ⋅ , ω − i )) .
Integrating over the realizations of the other DMs with respect to ( η ⊗ W ) ⊗ ( N − 1 ) (\eta\otimes\mathbb{W})^{\otimes(N-1)} ( η ⊗ W ) ⊗ ( N − 1 ) gives
J i ∗ ( p ; τ ˉ N , i , d ˉ N , i ) ≤ J i ∗ ( p ; τ , d ) . J_i^*(p;\bar{\tau}^{N,i},\bar{d}^{N,i})\leq J_i^*(p;\tau,d). J i ∗ ( p ; τ ˉ N , i , d ˉ N , i ) ≤ J i ∗ ( p ; τ , d ) .
Combining the above with (62) yields
J i N ( p ; τ ˉ N , i , d ˉ N , i ) ≤ J i N ( p ; τ , d ) + C N − 1 . J_i^N(p;\bar{\tau}^{N,i},\bar{d}^{N,i})\leq J_i^N(p;\tau,d)+\frac{C}{\sqrt{N-1}}. J i N ( p ; τ ˉ N , i , d ˉ N , i ) ≤ J i N ( p ; τ , d ) + N − 1 C .
and we conclude.
□ \square □