STABILITY OF TRAVELING WAVE SOLUTIONS IN A CREDIT RATING

MIGRATION FREE BOUNDARY PROBLEM

arXiv:2401.00198v1 [math.AP] 30 Dec 2023

CLAUDE-MICHEL BRAUNER, YUCHAO DONG, JIN LIANG, AND LUCA LORENZI

Abstract. In this paper, we study the stability of traveling wave solutions arising from a
credit rating migration problem with a free boundary, After some transformations, we turn
the Free Boundary Problem into a fully nonlinear parabolic problem on a fixed domain and
establish a rigorous stability analysis of the equilibrium in an exponentially weighted function
space. It implies the convergence of the discounted value of bonds that stands as an attenuated
traveling wave solution.

1. Introduction
1.1. Financial motivation. The evolution and globalization of the financial market have am-
plified the significance of credit risks, often elevating them to pivotal roles within the market
dynamics. This influence became unmistakably evident during the 2008 financial crisis and the
more recent wave of defaults among small banks in the United States. These instances serve as
poignant illustrations of the profound impact that credit risks can exert on the financial land-
scape. Consequently, the imperative of effectively managing credit risk has surged in importance
over time, underlining the critical need for robust risk mitigation strategies and practices.
Credit risks encompass two primary facets: default risk and credit rating migration risk. De-
fault risk signifies the likelihood that a company will fail to meet its debt obligations, while
credit rating migration risk pertains to alterations in a company’s credit rating triggered by
shifts in its financial condition, typically impacting the company’s equities and the potential
risks associated with related financial contracts. The quantification of credit risk through math-
ematical methodologies falls within two distinct frameworks: structural models and intensity
models. The intensity model posits that credit risk emanates from external factors, often mod-
eled using Markov chains, as referenced in [16]. In this context, the occurrences of default and/or
credit rating migrations hinge on an external transition intensity, as elucidated by Jarrow and
Longstaff in [10, 11]. However, in practical markets, it is evident that a company’s prevailing
financial condition significantly influences the probabilities of default and credit rating migra-
tions. To incorporate this endogenous factor, structural models emerge as pivotal tools for credit
risk modeling, with historical roots tracing back to Merton’s seminal work in 1974, as outlined
in [17]. Within such models, credit rating migrations and defaults are linked to the firm’s asset
value and its outstanding obligations.
In 2015, Liang and Zeng [13] first study the pricing problem of the corporate bond with credit
rating migration risk, where a predetermined migration threshold is given to divide asset value
into high and low rating regions, in which the asset value follows different stochastic processes.
Hu, Liang, and Wu [9] further develop this model, where the migration boundary is a free
boundary governed by a ratio of the firm’s asset value and debt. Furthermore, one contribution
of their work is to show that this problem admits an asymptotic traveling wave. Dong, Liang
and Brauner [8] further extend this framework to include default in the model. They also find
the asymptotic traveling wave. For some other works relative to this model, the readers can see
the survey paper [7]

2020 Mathematics Subject Classification. Primary: 35R35, 35B25; Secondary: 35C07, 91G20.
Key words and phrases. Free boundary problems; stability; traveling waves; fully nonlinear parabolic problems;
credit rating migration.
1
2 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

In previous research, the asymptotic behavior takes an important position and it also brings
up stability problem. Besides its own mathematical interest, the study of stability is also of
financial importance. As the traveling wave has a closed form, stability means the approximation
of the original solution by the traveling wave. In [9], the asymptotic traveling wave equation
offers insights into pricing a specific debt instrument subject to credit rating risk. This pricing
methodology hinges on evaluating the discounted value of a company’s assets. Given that [9]
focuses on a particular type of debt, a natural question is whether this pricing rule is suitable
for other kinds of long-term debts and robust when confronted with slight variations in model
coefficients. From a mathematical point of view, this question is closely related to the stability of
the traveling wave solutions. If stability can be established, it signifies the existence of a robust
and general pricing framework with credit risk, thereby expanding its utility beyond the specific
debt instrument initially considered in the study.

1.2. The model. Throughout the paper we will adopt the following notations as far as financial
parameters are concerned:
• δ > 0 is the risk discount rate;
• 0 < γ < 1 is the threshold proportion of the debt and value of the firm’s rating;
• r > 0 is the risk free interest rate;
• σH > 0 and σL > 0 represent the volatilities of the firm under the high and low credit
grades, respectively.
They verify the condition (see [12, (4.12)]),
1 2 1 2
σH < δ < σL , (1.1)
2 2
which hereafter defines the domain (A) of financially admissible parameters.
Our starting point is the Free Boundary Problem for the discounted value of the bond in
finance φ (see [12, Section 3])
 
∂φ 1 2 ∂2φ 1 2 ∂φ
(t, x) − σ (t, x) − r − σ (t, x) + rφ(t, x) = 0, (1.2)
∂t 2 ∂x2 2 ∂x
where x ∈ R is the spatial coordinate and t denotes the time. In the equation (1.2),
σ = σ(φ, x) = σH + (σL − σH )H(φ − γex−δt ),
where H is the Heaviside function.
The free boundary (or free interface) s(t) and the interface conditions are
  
   ∂φ
 φ(t, ·) = (t, ·) = 0,
x=s(t) ∂x x=s(t) (1.3)

 s(t)−δt
φ(t, s(t)) = γe ,
  + −
where f ξ=ξ0 = f (ξ0 ) − f (ξ0 ).
In order to establish convergence, as t tends to +∞, of φ(t, x) to a traveling wave solution, it
is convenient to make the following change of dependent variable (see [12, Section 3])
u(t, x) = ert φ(t, x). (1.4)
From (1.2) and (1.3) it follows that u satisfies the differential equation
 
∂u 1 ∂2u 1 ∂u
(t, x) − σ 2 2 (t, x) − r − σ 2 (t, x) = 0, (1.5)
∂t 2 ∂x 2 ∂x
with interface conditions at x = s(t):
  
   ∂u
 u(t) = (t) = 0,
x=s(t) ∂x x=s(t) (1.6)


u(t, s(t)) = γes(t)+(r−δ)t ,
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 3

In addition, the following boundary conditions hold at ±∞:

lim u(t, x) = 0, lim u(t, x) = 1. (1.7)
x→−∞ x→+∞

We look for a traveling wave solution K(x + ct) of system (1.5)–(1.7) which propagates at
velocity c = r − δ. In the moving coordinate (see [12, Section 3])
ξ = x + ct, η(t) = s(t) + ct,
the system for u = u(t, ξ) and free boundary η = η(t) reads
 
∂u 1 2 ∂2u 1 2 ∂u
(t, ξ) = σL 2 (t, ξ) + δ − σL (t, ξ), ξ < η(t),
∂t 2 ∂ξ 2 ∂ξ
  (1.8)
∂u 1 2 ∂2u 1 2 ∂u
(t, ξ) = σH (t, ξ) + δ − σ (t, ξ), ξ > η(t).
∂t 2 ∂ξ 2 2 H ∂ξ
From (1.6) we derive the following free interface conditions at ξ = η(t):
  
   ∂u
 u(t, ·) = (t, ·) = 0,
ξ=η(t) ∂ξ ξ=η(t) (1.9)


u(t, η(t)) = γeη(t) ,
together with the boundary conditions:
lim u(t, ξ) = 0, lim u(t, ξ) = 1. (1.10)
ξ→−∞ ξ→+∞

System (1.8)-(1.10) admits a unique steady solution (K, η ∗ ) which verifies the following system
(see [12, Lemma 4.7])
  
 1 2 d2 K 1 2 dK
 ξ < η∗ ,
 2 σL dξ 2 + δ − 2 σL dξ = 0,



  
 1
 2
 σ 2 d K + δ − 1 σ 2 dK = 0, ξ > η∗ ,
2 H dξ 2 2 H dξ

  

 dK ∗
 [K]ξ=η∗ =

 = 0, K(η ∗ ) = γeη ,

 dξ ξ=η∗

K(−∞) = 0, K(+∞) = 1,
which is defined by
 2 2

σL (2δ − σH )
η ∗ = ln 2 2
2δγ(σL − σH )
and
   

 η∗ 2δ ∗
 γe exp 1 − σ 2 (ξ − η ) ,
 ξ ≤ η∗ ,
L
K(ξ) =    (1.11)

 ∗ 2δ
 η ∗
 1 − (1 − γe ) exp 1 − 2 (ξ − η ) , ξ>η ∗
σH
The convergence of u(t, ξ) to K(ξ) as t tends to +∞ was proved in [12] in the case of a particular
initial condition. In this paper, we prove the nonlinear stability of the traveling wave K for all
the admissible values of parameters, using the techniques of [2–4, 6] developed for applications
in physics, especially in combustion theory. As we have already said, here stability indicates a
robust general rule of pricing under credit risk and our mathematical results confirm the validity
of the model of [12]. In other scientific fields, domains of instability are more likely to be identified
(see, e.g., [5]).
The paper is organized as follows: in Section 2, we fix the free interface η(t) to η ∗ . The new
spatial variable is denoted by y = ξ − η(t) + η ∗ . Then, we consider the perturbations
v(t, y) = u(t, y) − K(y), f (t) = η(t) − η ∗ .
4 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

According to the method of [3], we introduce the ansatz:

dK
v(t, y) = f (t) (y) + w(t, y)
dy
where w is the new unknown. After eliminating f and its derivative, we show that w verifies the
problem

 dw (t) = Lw(t) + F (w(t)),
dt (1.12)

B1 w(t) = G(w(t)).
The system for w does not contain the free interface which has been eliminated. It is a fully
nonlinear parabolic problem (see [15]) because the nonlinear term contains traces of second-order
derivatives with respect to the spatial variable y and, therefore, is of the same order as the linear
operator.
The analysis of the stability of the traveling wave solution K is now equivalent to the stability
of the trivial solution w = 0 of problem (1.12).
In Section 3, we prove that the realization L of the operator L in a suitable weighted space X
of continuous functions generates an analytic semigroup. As it is well known, the introduction
of exponentially weighted spaces for proving stability of traveling waves has been widely used
since the pioneering work of Sattinger (see [18]). The spectrum σ(L) consists of the union of
a halfline, strictly included in (−∞, 0), and the solutions of the dispersion relation Dλ = 0.
Section 4 is devoted to a thorough analysis of the dispersion relation. It turns out that, in the
domain of financially admissible parameters, the dispersion relation has no solutions. Therefore,
we establish the nonlinear stability of w = 0 in Section 5. The main result is Theorem 5.3.
In the last Section 6, we return to the problem for the discounted value of the bond φ =
e−rt u. We observe that, when t → +∞, φ(t, x) is asymptotically equivalent to the discounted or
attenuated traveling wave Φ(t, x) = e−rt K(x+ct). It is worth noting that attenuation of traveling
waves is a common phenomenon in physics due to loss of energy; for example, propagating seismic-
waves attenuate with time because Earth is not perfectly elastic (see [1]).
Notation. Throughout the paper, the derivatives of functions which depend on a single variable
will be denoted using a prime. For example, K ′′ will denote the second-order derivative of the
∂ ∂2
function K. Sometimes, we will find it convenient to write Dy and Dyy instead of ∂y and ∂y 2,

respectively. Moreover, for every interval I ⊂ R and every k ∈ N ∪ {0}, we denote by Cbk (I; C)
the set of all functions u : I → C which are continuously differentiable up to the order k in I
and are bounded together with all their derivatives. This space is endowed with its natural norm
Pk
kukCbk (I;C) = j=0 ku(j) k∞ where u(0) = u, and k · k∞ denotes the sup-norm over I.

2. The linearized model

2.1. The system with fixed interface. We set t′ = t, y = ξ − η(t) + η ∗ , i.e. y = x − s(t) + η ∗ .
When ξ = η(t), y = η ∗ , that is, the free interface is now at y = η ∗ .
In the coordinates (t′ , y), omitting the prime for simplicity, system (1.8)-(1.10) reads for y 6= η ∗ :
 
∂u ∂u 1 2 ∂2u 1 2 ∂u
− η̇ = σL + δ − σ , y < η∗ , (2.1)
∂t ∂y 2 ∂y 2 2 L ∂y
 
∂u ∂u 1 2 ∂2u 1 2 ∂u
− η̇ = σH + δ − σ , y > η∗ , (2.2)
∂t ∂y 2 ∂y 2 2 H ∂y
dη
where η̇ = . At the fixed interface y = η ∗ , the following interface conditions are satisfied:
dt
  
   ∂u
 u(t, ·) = (t, ·) = 0,
y=η ∗ ∂y y=η ∗


u(t, η ∗ ) = γeη(t) .
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 5

Next, we define the perturbations

v(t, y) = u(t, y) − K(y), f (t) = η(t) − η ∗ . (2.3)
and write the system verified by the pair (v, f ). From (2.1) and (2.2), we get the partial differential
equations:
   
∂v 1 2 ∂2v 1 2 ∂v ′ ∂v dK
= σL + δ − σ + f + , y < η∗ , (2.4)
∂t 2 ∂y 2 2 L ∂y ∂y dy
   
∂v 1 2 ∂2v 1 2 ∂v ′ ∂v dK
= σH + δ − σ + f + , y > η∗ , . (2.5)
∂t 2 ∂y 2 2 H ∂y ∂y dy
The following interface conditions are satisfied at y = η ∗ :
  
   ∂v
 v(t, ·) = (t, ·) = 0,
y=η ∗ ∂y y=η ∗ (2.6)

 ∗
v(t, η ∗ ) = γeη (ef (t) − 1).
It is readily seen that the term with f ′ contributes additional difficulties that we will overcome
in the next subsections.

2.2. Ansatz. According to the method of [3], we introduce the following ansatz:
v(t, y) = f (t)K ′ (y) + w(t, y), (2.7)
where w is the new unknown. We are going to derive a system verified by w.
Plugging (2.7) into (2.4)-(2.5), the system for (f, w) reads:
   
∂w 1 2 ∂2w 1 2 ∂w ′ ′′ ∂w
= σL 2 + δ − σL + f fK + , y < η∗ , (2.8)
∂t 2 ∂y 2 ∂y ∂y
   
∂w 1 2 ∂2w 1 2 ∂w ′ ′′ ∂w
= σH 2 + δ − σH + f fK + , y > η∗ . (2.9)
∂t 2 ∂y 2 ∂y ∂y
The interface conditions at y = η ∗ demand more attention, see [2, 6].
(i) As v(t, y) = f (t)K ′ (y) + w(t, y), it follows that
   
v(t, ·) y=η∗ = f (t)[K ′ ]y=η∗ + w(t, ·) y=η∗ .
Hence, [w(t, ·)]y=η∗ = 0.  
∗ ∗ 2
(ii) Since v(t, η ∗ ) = f (t)K ′ (η ∗ ) + w(t, η ∗ ) = γeη (ef (t) − 1) and K ′ (η ∗ ) = γeη 1 − 2 , it
δL
comes:
 
∗ ∗ 2δ
w(t, η ∗ ) = γeη (ef (t) − 1) − f (t)γeη 1− 2 . (2.10)
σL
(iii) Differentiating formula (2.7) for y 6= η ∗ and, then, taking the limit as y tends to η ∗ from
the left and from the right, it follows that
 
∂w
(t) = −f (t)[K ′′ ]y=η∗ .
∂y y=η ∗

Now, twice differentiating formula (1.11) and, then, taking the left- and right-limit at y = η ∗
gives so that
 2  2
′′ η∗ 2δ η∗ 2δ
[K ]y=η∗ = (γe − 1) 1 − 2 − γe 1− 2 . (2.11)
σH σL
Since
   
η∗ 2δ η∗ 2δ
(γe − 1) 1 − 2 = γe 1− 2 ,
σH σL
6 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

from formula (2.11) we deduce that

 
∗ 2δ 2δ
[K ′′ ]y=η∗ = γeη 1− 2 2
2 2
2 (σH − σL ).
σL σH σL
Hence,
   
∂w ′′ η∗ 2δ 2δ 2 2
(t, ·) = −f (t)[K ]y=η = f (t)γe
∗ 1− 2 2 σ 2 (σL − σH ). (2.12)
∂y y=η ∗ σL σH L

Remark 2.1. In contrast to u and v, w is not continuously differentiable at y = η ∗ (see [3]).

The next point is to establish a relation between f (t) and w(t, η ∗ ). According to formula
(2.10), let us define the function Φ : R → R, by setting
∗
 2δ 
Φ(x) = γeη ex − 1 − x + 2 x , x ∈ R.
σL
2δ ∗
Clearly, Φ vanishes at the origin and Φ′ (0) = γeη 2 . It is easy to show that the function Φ
σL  
is increasing in [x0 , +∞) and Φ(x0 ) < 0, where x0 = log 1 − σ2δ2 < 0 (note that 1 − σ2δ2 > 0
L L
due to condition (1.1)). Therefore, the restriction of the function Φ to the interval (x0 , +∞) is
invertible. Let us denote its inverse by Ψ, which is an increasing function defined on (Φ(x0 ), +∞)
and vanishes at the origin.
We may rewrite equation (2.10) in the more compact form w(t, η ∗ ) = Φ(f (t)). Therefore, is
w is small enough in some convenient norm, we can write f (t) in terms of w, via the formula
f (t) = Ψ(w(t, η ∗ )). (2.13)
Remark 2.2. Since Ψ is smooth in in (Φ(x0 ), +∞), using Taylor expansion centered at zero, we
can write
∗ σ
2
Ψ(w(t, η ∗ )) = Ψ′ (0)w(t, η ∗ ) + R(w(t, η ∗ )) = e−η L w(t, η ∗ ) + R(w(t, η ∗ )), (2.14)
2γδ
where R is the remainder.
2.3. The fully nonlinear problem. Thanks to equation (2.13) and Remark 2.2, f (t) can be
eliminated in the relation (2.12), obtaining the equivalent condition
     
∂w 2δ 1 2 2 ∗ ∗ η∗ 2δ 2δ 2 2
(t, ·) = 1− 2 2 (σL − σH )w(t, η ) + R(w(t, η ))γe 1 − 2 2 σ 2 (σL − σH )
∂y y=η ∗ σL σH σL σH L

The last step in the method consists in getting rid of f ′ in equations (2.8)-(2.9).
From (2.13), it comes:
∂w
f ′ (t) = Ψ′ (w(t, η ∗ )) (t, η ∗ ). (2.15)
∂t
∂w
To compute (t, η ∗ ), we take the trace of (2.8) at y = η−
∗
(or equivalently the trace of (2.9) at
∗
∂t
η+ ) and get
   
∂w ∗ 1 2 ∂2w ∗ 1 2 ∂w ∗ ′ ∗ ′′ ∗ ∂w ∗
(t, η ) = σL 2 (t, η− ) + δ − σL (t, η− ) + f (t) Ψ(w(t, η ))K (η− ) + (t, η− ) .
∂t 2 ∂y 2 ∂y ∂y
(2.16)
Let us multiply (2.16) by Ψ′ (w(t, η ∗ )). Taking (2.15) into account, we deduce that
2
 
1 2 ∂ w 1 2 ∂w
f ′ (t) = Ψ′ (w(t, η ∗ ))σL (t, η ∗
) + Ψ ′
(w(t, η ∗
)) δ − σ ∗
(t, η− )
2 ∂y 2 −
2 L ∂y
n ∂w o
+ Ψ′ (w(t, η ∗ ))f ′ (t) Ψ(w(t, η ∗ ))K ′′ (η− ∗
)+ ∗
(t, η− ) ,
∂y
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 7

or, equivalently,
 n o
′ ′ ∗ ∗ ′′ ∗ ∂w ∗
f (t) 1 − Ψ (w(t, η )) Ψ(w(t, η ))K (η− ) + (t, η− )
∂y
2
 
1 ′ ∗ 2 ∂ w ∗ ′ ∗ 1 2 ∂w ∗
= Ψ (w(t, η ))σL 2 (t, η− ) + Ψ (w(t, η )) δ − σL (t, η− ).
2 ∂y 2 ∂y
Assuming that w is small enough (in some norm), we get to the following formula for the velocity
of the perturbation of the free interface
1 ′ ∗ 2 ∂2 w ∗ ′ ∗ 1 2 ∂w

∗
′ 2 Ψ (w(t, η ))σL ∂y 2 (t, η− ) + Ψ (w(t, η )) δ − 2 σL ∂y (t, η− )
f (t) =  ∗ ) + ∂w (t, η ∗ )
. (2.17)
1 − Ψ′ (w(t, η ∗ )) Ψ(w(t, η ∗ ))K ′′ (η− ∂y −

Remark 2.3. Formula (2.17) is an implicit second-order Stefan condition, see [4].
Now, we define the linear differential operator L and the boundary operator B1 by setting
 2
 1 2 ∂ w


1 2 ∂w

 2 σL 2 + δ − 2 σL , y < η∗ ,
∂y ∂y
Lw = (2.18)

 1 ∂2w 1

∂w
 2 σH 2 2
+ δ − 2 σH , y>η ,∗
∂y 2 ∂y
and
   
∂w 2δ 1 2 2
B1 w = − Aw(η ∗ ), where A = 1 − 2 2 (σL − σH ) > 0. (2.19)
∂y y=η∗ σL σH
Using such operators, we can rewrite the fully nonlinear parabolic problem in the form:

 dw (t) = Lw(t) + F (w(t)),
dt (2.20)

B1 w(t) = G(w(t)),
where the nonlinear (quadratic) terms F and G are defined by
1 ′ ∗ 2 ∂2w ∗ ′ ∗


1 2 ∂w ∗
2 Ψ (w(t, η ))σL ∂y 2 (t, η− ) + Ψ (w(t, η )) δ − 2 σL ∂y (t, η− )
F (w(t, y)) =  ∗ ) + ∂w (t, η ∗ )
1 − Ψ′ (w(t, η ∗ )) Ψ(w(t, η ∗ ))K ′′ (η− ∂y −
 ∂w 
× Ψ(w(t, η ∗ ))K ′′ (y) + (t, y) , (2.21)
∂y
 
∗ 2δ 2δ
G(w(t)) = R(w(t, η ∗ ))γeη 1 − 2 2 2
2 σ 2 (σL − σH ),
σL σH L
 2
∗ 2δ
where K ′′ (η−
∗
) = γeη 1 − 2 .
σL
Remark 2.4. The problem for w does not contain the free interface which has been eliminated.
It is a fully nonlinear parabolic problem (see [15]) because the nonlinear term (2.21) contains
traces of second-order derivatives with respect to the spatial variable y and, therefore, it is of the
same order as the linear operator L.

3. Analysis of the linear problem

For notational convenience, we set
2δ 2δ
cH = 2 , cL = 2 .
σH σL
For financial motivations (see [12]), the domain (A) of admissible parameters is more specifically
cL < 1 < cH ≤ 3. (3.1)
In this section, we will prove that the realization of the operator L in a suitable weighted space
of continuous functions generates an analytic semigroup. We also characterize its spectrum. For
8 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

this purpose, we introduce the weights qH : [η ∗ , +∞) → R and qL : (−∞, η ∗ ] → R, defined as

follows:
 
1
qL (y) = exp − (cL − 1)y , y ∈ (−∞, η ∗ ],
2
 
1
qH (y) = exp − (cH − 1)y , y ∈ [η ∗ , +∞),
2
Definition 3.1. We denote by X the space of all the pairs w = (w1 , w2 ), where w1 : (−∞, η ∗ ] →
C and w2 : [η ∗ , +∞) → C are continuous functions such that
w1 (y) w2 (y)
kwkX := sup + sup < +∞
y≤η ∗ qL (y) y≥η ∗ qH (y)

We also introduce the subspace X2 of X of pairs w = (w1 , w2 ) such that w1 ∈ C 2 ((−∞, η ∗ ]; C)

and w2 ∈ C 2 ([η ∗ , +∞); C) and both Dy w := (Dy w1 , Dy w2 ) and Dyy w := (Dyy w1 , Dyy w2 ) belong
to X . It is normed by setting
kwkX2 := kwkX + kDy wkX + kDyy wkX .
Remark 3.2. From the definition, it turns out that if w belongs to X and cL < 1 < cH , then
w1 and w2 decrease to zero, with exponential rate, at −∞ and +∞, respectively.
Lemma 3.3. The space X2 can be equivalently characterized as the set of all pairs w = (w1 , w2 )
w1 w2
such that the functions and belong to Cb2 ((−∞, η ∗ ]; C) and Cb2 ([η ∗ , +∞); C), respec-
qL qH
tively. Moreover, the norm of X2 is equivalent to the classical norm of Cb2 ((−∞, η ∗ ]; C) ×
Cb2 ([η ∗ , +∞); C).
Proof. A simple computation shows that
 
w1 Dy w1 1 w1
Dy = + (cL − 1) ,
qL qL 2 qL
 
w1 Dyy w1 Dy w1 1 Dy w1
Dyy = + (cL − 1) + (cL − 1)2
qL qL qL 4 qL
w 1
in (−∞, η ∗ ]. Hence, belongs to Cb2 ((−∞, η ∗ ]; C) and
qL
w1 Dy w1 (y) 1 w1 (y) Dyy w1 (y)
≤ cL sup + (1 − cL )(3 − cL ) sup + sup .
qL C 2 ((−∞,η∗ ]) y≤η ∗ qL (y) 4 y≤η ∗ qL (y) y≤η ∗ qL (y)
b
w2
Similarly, qH belongs to Cb2 ([η ∗ , +∞); C) and
w2 Dy w2 (y) 1 w2 (y) Dyy w2 (y)
≤ cH sup + (c2H − 1) sup + sup .
qH Cb2 ([η ∗ ,+∞)) y≥η ∗ qH (y) 4 y≥η ∗ qH (y) y≥η ∗ qL (y)
Combining these last two estimates, we can infer that there exists a positive constant C,
independent of w such that
w1 w2
+ ≤ CkwkX2 .
qL C 2 ((−∞,η∗ ]) qH C 2 ([η∗ ,+∞)
b b

Vice versa, suppose that z1 ∈ Cb2 ((−∞, η ∗ ]; C) and z2 ∈ Cb2 ([η ∗ , +∞); C) and set w = (w1 , w2 ),
where w1 = z1 qL and w2 = z2 qH . Clearly, w belongs to X and kwkX = kz1 kCb ((−∞,η∗ ];C) +
kz2 kCb ([η∗ ,+∞);C) . Moreover, since
   
1 1 2
Dy w1 = Dy z1 − (cL − 1)z1 qL , Dyy w1 = Dyy z1 + (1 − cL )Dy z1 + (cL − 1) z1 qL
2 4
in (−∞, η ∗ ] and
   
1 1 2
Dy w2 = Dy z2 − (cH − 1)z2 qH , Dyy w2 = Dyy z2 − (cH − 1)Dy z2 + (cH − 1) z2 qH
2 4
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 9

in [η ∗ , +∞), it is an easy task to check that w ∈ X2 and

kwkX2 ≤ C2 kz1 kCb2 ((−∞,η∗ ];C) + kz2 kCb2 ([η∗ ,+∞);C)
for some positive constant C2 , independent of w. This completes the proof. 
3.1. The generation result. This the main result of this subsection.
Theorem 3.4. The realization L of the operator L in X , with domain
D(L) = {w ∈ X : Lw ∈ X , B(w) = 0} = {w ∈ X2 : B(w) = 0},
generates an analytic semigroup. Here, Bw = (B0 w, B1 w), where B0 w = w2 (η ∗ ) − w1 (η ∗ ) and
 B1 is defined
the operator n 2 by (2.19).2 Moreover,oiits spectrum σ(L) consists of the union of the
c c
halfline −∞, − min 8L (c2L − 1)2 , 8H (c2H − 1)2 and the solutions of the dispersion relation
Dλ = 0, where
Dλ = A − µH L
− + µ+ ,
and
s
1 1 2λ
µJ± = − (cJ − 1) ± (cJ − 1)2 + 2 , J ∈ {H, L}.
2 4 σJ
Proof. Let us fix f ∈ X , λ ∈ C and consider the resolvent equation λw − L w = f . We need to
distinguish some cases. i  i
c2 c2
/ −∞, − 8L (c2L − 1)2 ∪ −∞, − 8H (c2H − 1)2 , then an easy computation reveals
Case 1. If λ ∈
that the more general solution wλ of such an equation, which belongs to the space X , is given
by
 Z η∗  Z y 
1 −µL L
+ s f (s)ds eµ+ y +
1 −µL L
− s f (s)ds eµ− y
w1,λ (y) = c1,λ + L L
e 1 L L
e 1
µ+ − µ− y µ+ − µ− −∞
(3.2)
for every y ∈ (−∞, η ∗ ] and
 Z +∞   Z y 
1 −µH s µH y 1 −µH s H
w2,λ (y) = H H
e + f2 (s)ds e + + d1,λ + H H
e − f2 (s)ds eµ− y
µ+ − µ− y µ+ − µ− η ∗

(3.3)
for every y ∈ [η ∗ , +∞).
Let us now impose the boundary conditions at η ∗ . The first condition is B0 w1,λ = 0, i.e.,
w1,λ (η ∗ ) = w2,λ (η ∗ ), which means that
L ∗ Z η∗ H ∗ Z +∞
µL η∗ µH η∗ eµ− η −µL s eµ+ η H
c1,λ e + − d1,λ e − =− L L
e − f1 (s)ds + H H
e−µ+ s f2 (s)ds, (3.4)
µ+ − µ− −∞ µ+ − µ− η ∗
∂w2 ∗ ∂w1 ∗
whereas the second condition B1 wλ = 0, i.e., (η )− (η )−Aw2 (η ∗ ) = 0, reads as follows:
∂y ∂y
L ∗ H ∗
µL
+e
µ+ η
c1,λ + (A − µH− )e
µ− η
d1,λ
 Z +∞   Z η∗ 
µH − A −µH H ∗ µL −µL L ∗
= H+ e + s f (s)ds eµ+ η −
2
−
e − s f (s)ds eµ− η .
1 (3.5)
µ+ − µH − η ∗ µ L − µL
+ − −∞

The system (3.4)-(3.5) is uniquely solvable if and only if and only if Dλ 6= 0, i.e., if and only if
 2  s s
σL 1 1 2λ 1 2λ
(1 − cL ) 2 − 1 + (cH + cL ) − 1 + (cH − 1)2 + 2 − (cL − 1)2 + 2 6= 0.
σH 2 4 σH 4 σL
In this case,
Z η∗  Z +∞ 
A − µH + µL −µL (µL L ∗ 1 −µH H L ∗
c1,λ =− L −L − e −s f1 (s)ds e − −µ+ )η + e +s f2 (s)ds e(µ+ −µ+ )η
(µ+ − µ− )Dλ −∞ Dλ η∗
10 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

and
Z η∗   Z +∞ 
1 L L H ∗ A − µH + µL H H H ∗
d1,λ = e−µ− s f1 (s)ds e(µ− −µ− )η − H + H + e−µ+ s f2 (s)ds e(µ+ −µ− )η ,
Dλ −∞ (µ+ − µ− )Dλ η∗

On the other hand, if Dλ = 0, then the equation λw − Lw = 0, subject to the boundary

conditions Bw = 0, admits more than a unique solution, i.e., every solution to the dispersion
relation Dλ = 0 lies in thepoint spectrum of the
i operator (L, B).
c2L 2 2
Suppose now that λ ∈ −∞, − 8 (cL − 1) . We claim that λ belongs to the spectrum of the
operator L. For this purpose, we observe that, if we take f = 0, then the more general solution
to the equation λw − Lw = f , which belongs to X , has the first component which is given by
L L
w1 (y) = c1 eµ+ y + c2 eµ− y , y ∈ (−∞, η ∗ ],
whereas the second component is given by
L
w2 (y) = d1 eµ− y , y ∈ (−∞, η ∗ ],
 i
c2
/ −∞, − 8H (c2H − 1)2 , and
if λ ∈
L L
w2 (y) = d1 eµ− y + d2 eµ+ y , y ∈ (−∞, η ∗ ],
 i
c2
if λ ∈ −∞, − 8H (c2H − 1)2 . Therefore, we have at least three parameters to identify and two
 i
c2
boundary conditions. Clearly, this is not possible. Hence, the interval −∞, − 8L (c2L − 1)2 is
contained in the spectrum of the operator L.  i
c2
A completely similar argument shows that also the interval −∞, − 8H (c2H − 1)2 is contained
in σ(L).
Now, we observe that the function wλ , defined by (3.2) and (3.3), has first- and second-order
derivatives which belong to X . Indeed,
 Z η∗  Z y 
∂w1,λ L 1 −µL s µL y µL
− −µL s L
(y) =µ+ c1,λ + L e + f1 (s)ds e + + L e − f1 (s)ds eµ− y
∂y µ+ − µL − y µ + − µ L
− −∞

for every y ∈ (−∞, η ∗ ] and

 Z +∞   Z y 
∂w2,λ µH H H 1 −µH H
(y) = H + H e−µ+ s f2 (s)ds eµ+ y + µH
− d1,λ + H H
e − s f (s)ds eµ− y
2
∂y µ+ − µ− y µ+ − µ− η ∗
for every y ∈ [η ∗ , +∞). These two functions are clearly continuous in (−∞, η ∗ ] and in [η ∗ , +∞),
w1,λ w2,λ
respectively. Moreover, it is an easy task to show that the functions and are bounded
qL qH
in (−∞, η ∗ ] and in [η ∗ , +∞), respectively. Since
   
∂ 2 w1,λ 2 1 2
= 2 λw1,λ − δ − σL Dy w1,λ (y) − f1 (y) , y ∈ (−∞, η ∗ ],
∂y 2 σL 2
   
∂ 2 w2,λ 2 1 2
= 2 λw2,λ − δ − σH Dy w2,λ (y) − f1 (y) , y ∈ [η ∗ , +∞)
∂y 2 σH 2
and wλ , Dy wλ and f belong to X , by difference we conclude that Dyy w ∈ X .
The characterization of the domain of L and of its spectrum is complete.
To conclude the proof, we need to show that L is a sectorial operator. For this purpose, we
set R(λ, L)f := wλ for every λ ∈ ρ(L) and show that there exist two positive constants C and
M such that
kR(λ, L)f kL(X ) ≤ M |λ|−1 kf kX , (3.6)
for every λ ∈ C, with Reλ ≥ M . In view of [14, Proposition 3.2.8] this will be enough to conclude
that the operator L is sectorial in H .
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 11

We begin by estimating the function w1,1,λ : (−∞, η ∗ ] → C, defined by

 Z η∗  Z y 
1 −µL L
+ s f (s)ds eµ+ y +
1 −µL L
− s f (s)ds eµ− y
w1,1,λ (y) = L e 1 e 1
µ+ − µL− y µ L − µL
+ − −∞

for every y ∈ (−∞, η ∗ ]. Observe that

Z η∗ Z y 
w1,1,λ (y) 1 f1 (s) 1 L L 1 L L
≤ L sup e 2 (µ+ −µ− )(y−s) ds + e− 2 (µ+ −µ− )(y−s) ds
qL (y) |µ+ − µL
− | s≤η qL (s) y −∞
Z η ∗ −y Z +∞ 
2 f1 (s) − 12 Re(µL L
+ −µ− )s
1 L L
≤ sup e ds + e− 2 Re(µ+ −µ− )s ds
|µL
+ − µ L|
− s≤η qL (s) 0 0
2 1 f1 (s) 1 L L ∗

= sup 2 − e− 2 Re(µ+ −µ− )(η −y)
|µL
+
L L L
− µ− | Re(µ+ − µ− ) s≤η qL (s)
4 1 f1 (s)
≤ sup
|µL
+
L L L
− µ− | Re(µ+ − µ− ) s≤η L (s)
q
for every y ∈ (−∞, η ∗ ]. Since
s
8λ
µL
+ − µL
− = (cL − 1)2 + 2
σL
it follows easily that
s s
p (cL − 1)2 8 p 8 (cL − 1)2 2 p
|µL L
+ − µ− | = |λ| + 2 ≥ |λ| 2 − ≥ |λ|,
λ σL σL |λ| σL
σ2 (c −1)2
provided that |λ| ≥ L L4 .
√
Moreover, recalling that by · we mean the main determination of the square root in C and
√ √
2
p
Re z = 2 |z| + Re(z) for every z ∈ C, we can also estimate
√ s √ s
L L 2 8λ 8Re(λ) 2 8λ
Re(µ+ − µ− ) ≥ (cL − 1)2 + 2 + (cL − 1)2 + 2 ≥ (cL − 1)2 + 2 , (3.7)
2 σL σL 2 σL
σ2 (c −1)2
if the real part of λ is nonnegative. Summing up, if Re(λ) ≥ L L4 , then
√
w1,1,λ (y) 2σL f1 (s)
sup ≤ sup .
y≤η ∗ qL (y) |λ| s≤η qL (s)

In a completely similar way, we can estimate the function w2,1,λ : [η ∗ , +∞) → C, defined by
 Z +∞  Z y 
1 −µH H
+ s f (s)ds eµ+ y +
1 −µH H
− s f (s)ds eµ− y
w2,1,λ (y) = H e 2 e 2
µ+ − µH− y µ H − µH
+ − η ∗

for every y ∈ [η ∗ , +∞). It turns out that

√
w2,1,λ (y) 2σH f2 (s)
sup ≤ sup .
y≤η ∗ qH (y) |λ| s≥η qL (s)
L
We now turn our attention to the function y 7→ w1,2,λ (y) = c1,λ eµ+ y and note that
w1,2,λ (y) 1 L L
≤ |c1,λ |e 2 Re(µ+ −µ− )η∗ , y ∈ (−∞, η ∗ ]. (3.8)
qL (y)
Hence, we need to estimate the constant c1,λ . For this purpose, we observe that
Z η∗
|A − µH + µL | L L L ∗
|c1,λ | ≤ L −L − e−µ− s f1 (s)ds eRe(µ− −µ+ )η
|µ+ − µ− ||Dλ | −∞
Z +∞
1 H H L ∗
+ e−µ+ s f2 (s)ds eRe(µ+ −µ+ )η
|Dλ | η∗
12 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

 Z η∗ 
σL |A − µH L
− + µ− | f1 (s) 1
Re(µ L L
)s L L ∗
≤ p sup e2 + −µ− ds eRe(µ− −µ+ )η
2 |λ||Dλ | s≤η qL (s) −∞
 Z +∞ 
1 f2 (s) − 12 Re(µH −µH )s H L ∗
+ sup e + − ds eRe(µ+ −µ+ )η
|Dλ | s≥η qH (s) η ∗

σL |A − µH + µL −| 1 f1 (s) − 21 Re(µL+ −µL+ )η∗

≤ p − L L
sup e
|λ||Dλ | Re(µ+ − µ− ) s≤η qL (s)
2 1 f2 (s) 1 Re(µH H L ∗
+ +µ− −2µ+ )η .
+ sup e2 (3.9)
|Dλ | Re(µH+ − µ H)
− s≥η qH (s)
Note that
Dλ =A − µH− + µ+
L

  s s
2δ 1 2 2 1 1 2λ 1 2λ
= 1− 2 2 (σL − σH ) + 2 (cH − cL ) + (cH − 1)2 + 2 + (cL − 1)2 + 2
σL σH 4 σH 4 σL
  s s
cH 1 1 2
2λ 1 2λ
=(1 − cL ) − 1 + (cH − cL ) + (cH − 1) + 2 + (cL − 1)2 + 2
cL 2 4 σH 4 σL
s s
cH 1 1 2λ 1 2λ
= + (cL − cH ) − 1 + (cH − 1)2 + 2 + (cL − 1)2 + 2 ,
cL 2 4 σH 4 σL
so that
s s
p 1 2 1 2 cH 1
|Dλ | ≥ |λ| (cH − 1)2 + 2 + (cL − 1)2 + 2 − + (cL − cH ) − 1
4λ σH 4λ σL cL 2
p
≥C1 |λ| (3.10)
for a positive constant C1 and every λ ∈ C with sufficiently large real part. In a completely
similar way, we can check that there exists a positive constant C2 such that
p
|A − µH L
− + µ− | ≥ C2 |λ| (3.11)
for every λ ∈ C with sufficiently large real part.
From (3.9)-(3.11) we deduce that
 
C3 f1 (s) − 1 Re(µL+ −µL− )η∗ f1 (s) 1 Re(µH H L
+ +µ− −2µ+ )η
∗
|c1,λ | ≤ sup e 2 + sup e2
|λ| s≤η qL (s) s≤η qL (s)

for every λ ∈ C, with sufficiently large real part, and some positive constant C3 , independent of
λ, and then that
 
w1,2,λ (y) C3 f1 (s) f1 (s) 1 Re(µH +µH −µL −µL )η ∗
≤ sup + sup e 2 + − + =
qL (y) |λ| s≤η qL (s) s≤η qL (s)
 
C3 f1 (s) f1 (s) (cH −cL )η∗
= sup + sup e (3.12)
|λ| s≤η qL (s) s≤η qL (s)

for every y ∈ (−∞, η ∗ ] and the same values of λ as above. From (3.7) and (3.12) it follows that
w1,λ (y) C4
sup ≤ kf kX (3.13)
y≤η ∗ qL (y) |λ|
for every λ ∈ C with Reλ ≥ M1 . Here, C4 and M1 are positive constants, independent of λ.
H
To complete the estimate of the function w2,λ , we need to consider the term d1,λ eµ− y . For
this purpose, we observe that
 Z η∗ 
1 f1 (s) 1
Re(µL −µL )s L L ∗
|d1,λ | ≤ sup e 2 + − ds eRe(µ− −µ+ )η
|Dλ | s≤η∗ qL (s) −∞
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 13

Z +∞ 
|A − µH + µL | f2 (s) − 12 Re(µH H H H ∗
+ H −H + sup e + −µ− )s ds eRe(µ+ −µ− )η
(µ+ − µ− )|Dλ | s≥η∗ qH (s) η∗
2 1 f1 (s) − 21 Re(µL+ −µL− )η∗
≤ L L
sup e
|Dλ | Re(µ+ − µ− ) s≤η∗ qL (s)
|A − µH L
− + µ+ | 2 f2 (s) 1 Re(muH H ∗
+ sup e2 + −µ− )η .
(µ+ − µ− )|Dλ | Re(µH
H H
+ − µ H)
− s≥η ∗ qH (s)
It follows that
1 H H 2 1 f1 (s) 1 Re(µL+ +µL− −µH H ∗
|d1,λ |e− 2 Re(µ+ −µ− )y ≤ sup e2 + −µ− )η
|Dλ | Re(µL
+ − µ L)
− s≤η ∗ qL (s)
|A − µH L
− + µ+ | 2 f2 (s)
+ H H H H
sup
(µ+ − µ− )|Dλ | Re(µ+ − µ− ) s≥η∗ qH (s)
C5
≤ kf kX (3.14)
|λ|
for every y ∈ [η ∗ , +∞) and λ ∈ C, with sufficiently large real part. From (3.8) and (3.14) we
deduce that there exists a positive constant C6 such that
f2 (y)
sup ≤ C6 kf kX (3.15)
y≥η ∗ qH (y)
for every λ ∈ C with Reλ ≥ M2 and some positive constants C6 , independent of λ, and M2 .
From (3.13) and (3.15), estimate (3.6) follows at once, with M = max{M1 , M2 }. The proof is
now complete. 

For further use, the following corollary will be particularly useful.

Corollary 3.5. For every λ ∈ ρ(L), every f ∈ X , every g ∈ R, the Cauchy problem

λw − Lw = f ,
(3.16)
Bw = (0, a),
admits a unique solution w ∈ X2 . Moreover, there exists a positive constant C, independent of
the data and of λ, such that
p
|λ|kwkX ≤ C(kf k + |λ||a|) (3.17)
for λ ∈ C with sufficiently large real part.
Proof. In view of the results in Theorem 3.4, it suffices to deal with the Cauchy problem

λw − Lw = 0,
Bw = (0, a).
Adapting the arguments used in the quoted theorem, it can be easily checked that the solution
to the previous equation is given by w = (w1 , w2 ), where
a µ+ ∗
w1 (y) = e L (y−η ) , y ∈ (−∞, η ∗ ],
Dλ
a µ− (y−η∗ )
w2 (y) = e H , y ∈ [η ∗ , +∞).
Dλ
1
Clearly, w ∈ X2 and taking (3.10) into account, we can show that kwkX ≤ |λ|− 2 |a|.
Estimate (3.17) now follows from the previous estimate and (3.6), observing that the solution
to the Cauchy problem (3.16) is given by R(λ, L) + w. 

We conclude this section by characterizing some interpolation spaces between X and D(L).
For this purpose, we introduce the following definition.
14 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

Definition 3.6. For every α ∈ (0, 1), we denote by Xα , the set of all functions w ∈ X such that
w1 w2
the functions and are α-Hölder continuous in (−∞, η ∗ ] and in [η ∗ , +∞), respectively. It
qL qH
is normed, by setting
w1 w2
kwkXα = +
qL C α ((−∞,η∗ ]) qH C α ([η∗ ,+∞))
b b

for every w ∈ Xα .
We also introduce the space X2+α , defined as the set of all functions w ∈ X2 such that
Dyy w = (Dyy w1 , Dyy w2 ) ∈ Xα . It is normed by setting
   
Dyy w1 Dyy w2
kwkX2+α = kwkX2 + +
qL C α ((−∞,η ∗ ]) qH C α ([η ∗ ,+∞))
b b

for every w ∈ X2+α .

Theorem 3.7. For every α ∈ (0, 1) it holds that
DL (α/2, ∞) = {w ∈ Xα : B0 w = 0}, DL (1 + α/2, ∞) = {w ∈ X2 : Bw = 0, B0 Lw = 0},
with equivalence of the corresponding norms. Moreover, {w ∈ X1 : B0 w = 0} ֒→ DL (1/2, ∞).
Proof. Throughout the proof, C denotes a positive constant, independent of t, λ and the functions
that we will consider, which may vary from line to line.
Let us fix a function w ∈ Xα (here, α ∈ (0, 1]). Since w1 (η ∗ ) = w2 (η ∗ ), the function ζ : R → R,
defined by

w1 (y), y ≤ η ∗ ,
ζ(y) =
w2 (y), y > η ∗
is α-Hölder continuous in R. Moreover, its Cbα (R)-norm coincides with the Xα -norm of w.
We regularize function ζ by convolution. For this purpose, let ϕ ∈ Cc∞ (R) be a nonnegative
function with L1 -norm over R equal to one. For every t > 0, we set
Z
ζt (x) = t ϕ(ty)ζ(x − y)dy, x ∈ R.
R
It is an easy task to check that
kζt − ζk∞ ≤ Ct−α kwkXα , (3.18)
1−α
kDx ζt k∞ ≤ Ct kwkXα , (3.19)
kDxx ζt k∞ ≤ Ct2−α kwkXα (3.20)
for every t > 0.
Now, we are ready to prove that w ∈ DL (α/2, ∞). For this purpose, we recall that DL (α/2, ∞)
can be characterized as the set of all functions w ∈ X for which there exists λ0 > 0 such that
α
[w]DL (α/2,∞) := sup λ 2 kLR(λ, L)wkX < +∞
λ≥λ0

and the norm of DL (α/2, ∞) is equivalent to the norm k · kX + [ · ]DL (α/2,∞) .

We introduce the function ζ√λ , whose components are the restrictions of ζ√λ to (−∞, η ∗ ] and
to [η ∗ , +∞), respectively. Next, we split,
LR(λ, L)w = LR(λ, L)(w − ζ√λ ) + LR(λ, L)ζ√λ ,
so that
kLR(λ, L)wkX ≤kLR(λ, L)(w − ζ√λ )kX + +kLR(λ, L)ζ√λ kX
α
≤Cλ− 2 kLR(λ, L)kL(X ) kwkXα + kLR(λ, L)ζ√λ kX . (3.21)
To estimate the second term in the last side of the previous inequality, we rewrite the function
uλ := LR(λ, L)ζ√λ in the following form:
uλ = (L − λ + λ)R(λ, L)ζ√λ = −ζ√λ + λR(λ, L)ζ√λ ,
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 15

so that, applying operator L to the first and the last side of the previous formula, we deduce
that uλ is a solution to the equation λuλ − Luλ = −Lζ√λ . Moreover,
Buλ = −Bζ√λ + λBR(λ, L)ζ√λ = −Bζ√λ .
Using (3.19) and (3.20), we can easily infer that
α
kLζ√λ kX ≤ Cλ1− 2 kwkXα
and
1 α
kBζ√λ kX ≤ Cλ 2 − 2 kwkXα
for every λ ≥ 1.
Now, applying estimate (3.17), we deduce that, if λ is sufficiently large, then
p α
λkuλ kX ≤ C(kLw√λ kX + |λ|kBζ√λ kX ) ≤ C|λ|1− 2 kζkXα
so that
α
kLR(λ, L)ζ√λ kX = kuλ kX ≤ C|λ|− 2 kwkXα
This estimate and (3.21) show that there exists λ0 > 0 such that
α
λ 2 kLR(λ, L)w√λ kX ≤ CkwkXα
for every λ ≥ λ0 . The embedding {w ∈ Xα : B0 w = 0} ֒→ DL (α, ∞) follows for every α ∈ (0, 1].
Let us now fix α ∈ (0, 1) and prove that DL (α/2, ∞) is continuously embedded into {w ∈
Xα : B0 w = 0}. For this purpose, we recall that DL (α/2, ∞) = (X , D(L))α/2,∞ with equiva-
lence of the corresponding norms. Since D(L) is continuously embedded into X2 , it follows that
DL (α/2, ∞) ֒→ (X , X2 )α/2,∞ . To characterize this latter space, we take advantage of Lemma
∗ ∗
 the isomorphism T : X → Cb ((−∞, η ]; C) × Cb ([η , +∞); C), defined by
3.3 and introduce
Tw = w w2
qL , qH
1
for every w ∈ X , which is also bijective between X2 and Cb2 ((−∞, η ∗ ]; C) ×
Cb ([η ∗ , +∞); C). The theory of interpolation spaces, it follows that T is also bijective between
(X , X2 )α,∞ and (Cb ((−∞, η ∗ ]; C)×Cb ([η ∗ , +∞); C), Cb2 ((−∞, η ∗ ]; C)×Cb2 ([η ∗ , +∞); C))α/2,∞ =
(Cb ((−∞, η ∗ ]; C); Cb2 ((−∞, η ∗ ]; C))α/2,∞ × (Cb ([η ∗ , +∞); C), Cb2 ([η ∗ , +∞); C))α/2,∞ , with equiv-
alence of the corresponding norms. Since this latter space coincides with Cbα ((−∞, η ∗ ]; C) ×
Cbα ([η ∗ , +∞); C), i.e., with the space Xα , with equivalence of the corresponding norms. We have
so proved that DL (α/2, +∞) is continuously embedded into Xα . To complete the proof of the
inclusion DL (α/2, +∞) ֒→ {w ∈ Xα : B0 w}, we recall that D(L) is dense in DL (α/2, +∞), with
respect to the norm of this latter space. Thus, we can determine a sequence {vn } ⊂ D(L) such
that kvn − wkDL (α/2,+∞) converges to 0 as n tends to +∞. In particular, kvn − wkX vanishes
as n tends to +∞. Since B0 vn = 0 for every n ∈ N, letting n tend to +∞, we get the condition
B0 w = 0.
The topological equivalence DL (α/2, +∞) = {w ∈ Xα : B0 w = 0} is now checked.
Let us now prove the second characterization, For this purpose, we begin by recalling that
DL (1 + α/2, +∞) is the set of all w ∈ D(L) such that Lw ∈ Xα . From Theorem 3.4 it follows
that D(L) ⊂ X2 and
kwkD(L) ≤ C1 kwkX2 , w ∈ D(L). (3.22)
On the other hand, since
     
1 2 ∂ 2 w1 1 2 ∂w1 1 2 ∂ 2 w2 1 2 ∂w2
Lw = σ + δ − σL , σ + δ − σH , w ∈ D(L),
2 L ∂y 2 2 ∂y 2 H ∂y 2 2 ∂y
it follows that    
∂ 2 wj 2 1 2 ∂wj
= (Lw)j − δ − σ , j = 1, 2, (3.23)
∂y 2 σL2 2 L ∂y
so that  
Dyy w1 Dy w1
≤ C kLwkX + . (3.24)
qL ∞ qL ∞
16 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

Further, we recall that the space Cb1 ((−∞, η ∗ ]) belongs to the class J1/2 between Cb ((−∞, η ∗ ])
and Cb2 ((−∞, η ∗ ]), which means that
1 1
kf kCb1 ((−∞,η∗ ]) ≤ Ckf k∞
2
kf kC2 2 ((−∞,η∗ ]) , f ∈ Cb2 ((−∞, η ∗ ]).
b
w1
Applying this estimate with f = , it follows that
qL
1   12
Dy w1 w1 w1 2 w1 Dy w1 Dyy w1
−C ≤C + + .
qL ∞ qL ∞ qL ∞ qL ∞ qL ∞ qL ∞
Therefore, using Cauchy-Schwarz inequality, from the previous inequality we deduce that
Dy w1 w1 Dyy w1 w1 Dy w1
−C ≤ε + Kε + Cε
qL ∞ qL ∞ qL ∞ qL ∞ qL ∞
for every ε > 0 and some positive constant Kε , independent of w and blowing up as ε tends to
0. Taking ε sufficiently small, we conclude that
Dy w1 Dyy w1 w1
≤ε + Kε′ (3.25)
qL ∞ qL ∞ qL ∞
for some positive constant Kε′ , independent of w1 . Replacing this inequality into (3.24), gives
Dyy w1 Dyy w1 w1
≤ CkLwkX + Cε + CKε′ .
qL ∞ qL ∞ qL ∞
Taking ε smaller, if needed, we can get rid of the norm of Dyy w1 from the right-hand side of
the previous estimate and conclude that
 
Dyy w1 w1
≤ C kLwkX + ≤ CkwkD(L) . (3.26)
qL ∞ qL ∞
From (3.25) and (3.26), it follows that
w1 Dy w1 Dyy w1
+ + ≤ CkwkD(L) .
qL ∞ qL ∞ qL ∞
Arguing in the same way, we can deal with the function w2 and we conclude that
kwkX2 ≤ CkwkD(L) , w ∈ D(L). (3.27)
Estimates (3.22) and (3.27) show that the graph norm of L is equivalent to the norm of X2
on D(L).
We can now complete the characterization of DL (1 + α/2, +∞). For this purpose, we fix
w ∈ DL (1 + α/2, +∞). Then, Lw ∈ Xα , which in view of (3.23), implies that Dyy w ∈ Xα , so
that w ∈ X2+α . Moreover,
    !
Dyy w1 2 1 2 Dy w1
≤ 2 kLwkXα + δ − σL
qL C α (−∞,η ∗ ]) σL 2 qL C α ((−∞,η∗ ])
!
1 2 w1
≤C kLwkXα + + δ − σL
2 qL C 1 ((−∞,η∗ ])
b

≤C(kLwkXα + kwkX2 )
≤CkwkDL (1+α/2,+∞)
for every w ∈ DL (1 + α/2, +∞). Clearly, Bw = 0 since w ∈ D(L), Moreover, B0 Lw = 0 since
Lw ∈ DL (α/2, +∞). We have so proved the topological inclusion DL (1 + α/2, +∞) ֒→ X2+α .
Viceversa, let us suppose that w ∈ X2+α is such that Bw = 0 and B0 Lw = 0. Then, w ∈ D(L)
and Lw ∈ X2+α . Since B0 Lw = 0, from the characterization of the space DL (α/2, +∞), it follows
that Lw ∈ DL (α/2, +∞). Therefore, w ∈ DL (1 + α/2, +∞). Moreover, kLwkXα ≤ CkwkX2+α .
From this estimate and the previous results, we conclude that X2+α ֒→ DL (1 + α/2, +∞). The
proof is now complete. 
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 17

4. A thorough analysis of the dispersion relation

In this subsection, we analyze the space the solution to the dispersione relation Dλ in order
to have a complete characterization of the spectrum of the operator L and its localization with
respect to the imaginary axis.
Theorem 4.1. The solutions of the dispersion relation Dλ = 0 are contained in the left-halfplane
{λ ∈ C : Reλ > 0} if and only if


 (cL − 2)(cH − cL ) ≥ 0,


(cL − 1)(2c2H + cH c2L − 4cH cL + c2L ) ≤ 0,

 (cL − cH )(−σH2
cH c2L + 2σH2 2
cH cL − 2σH 2 2
cH − σH 2
cL + 2σH 2
cL + 2σL 2
cH cL − 2σL cH

 2 3 2 2 2
+σL cL − 3σL cL + 2σL cL ) ≥ 0
or


 (cL − 2)(cH − cL ) ≥ 0,



 (cL − 1)(2c2H + cH c2L − 4cH cL + c2L )≤0,


 2
(cL − cH )(−σH cH c2L + 2σH2
cH cL − 2σH 2 2 2
cH − σH 2
cL + 2σH 2
cL + 2σL 2
cH cL − 2σL cH
2 3 2 2 2

 +σ L c L − 3σL c L + 2σ L c L )<0



 cH (cL − 1)2 (cH − cL )2 (cH + c2L − 2cL )≤0,



∆ > 0,
where ∆ is defined in (4.6).
Proof. We recall that the function λ 7→ Dλ is defined by
s s
cH 1 1 2
2λ 1 2λ
Dλ = + (cL − cH ) − 1 + (cH − 1) + 2 + (cL − 1)2 + 2
cL 2 4 σH 4 σL

for every λ ∈ C. The equation Dλ = 0 is easily seen to be equivalent to

s s
1 2
2λ 1 2λ cH 1
(cH − 1) + 2 + (cL − 1)2 + 2 = 1 − − (cL − cH )
4 σH 4 σL cL 2

Clearly, if 1 − ccHL − 12 (cL − cH ) < 0 then the previous equation does not admit solutions in C
√
since the left-hand side of the above equation has nonnegative real part (we recall that by · we
denote the principal square root).
Hence, from now on, we assume that
cH 1
1− − (cL − cH ) ≥ 0, i.e., (cL − 2)(cH − cL ) ≥ 0. (4.1)
cL 2
Then, squaring the above equation, we find the equivalent equation
s s
1 2λ 1 2λ
2 (cH − 1)2 + 2 (cL − 1)2 + 2
4 σH 4 σL
 
c2 1 1 cH 5 c2H 1 1 1
= H − c H c L + − 2 + c H − − c L − 2λ + 2 .
c2L 2 2 cL 2 cL 2 2
σH σL
Clearly, the previous equation cannot admit solutions λ with
2 2
 2 
σH σL cH 1 1 cH 5 c2H 1
Re(λ) > cH,L := 2 + σ2 ) − cH cL + − 2 + cH − − cL
2(σH L c2L 2 2 cL 2 cL 2
2 2
σH σL (cL − 1)(2c2H + cH c2L − 4cH cL + c2L )
=− 2 + σ 2 )c2 . (4.2)
4(σH L L
18 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

In particular, from (4.2) it follows that the dispersion relation Dλ = 0 cannot admit solution
with nonnegative real parts if (cL − 1)(2c2H + cH c2L − 4cH cL + c2L ) > 0. Hence, from now on we
assume that
(cL − 1)(2c2H + cH c2L − 4cH cL + c2L ) ≤ 0. (4.3)
On the other hand, if Re(λ) ≤ cH,L , then we can square once more the above equation and
get
1 2 2 1 1 1 1 1 1 1
c cL + c2H − c2H cL + c2L + − cL − cH c2L − cH + cH cL
4 H  4 2 4 4 2 2 2
1 2 1 2 16 2
+ 2λ 2 (cH + 1 − 2cH ) + 2 (cL + 1 − 2cL ) + 2 2 λ
σL σH σ σL
  
8λ 1 2 1 1 2λ
= c2H + 1 − 2cH + 2 c + − cL + 2
σH 4 L 4 2 σL
 2
c4 1 2 2 1 c2H 37 2 c4H 1 2 2 1 1 c3 c3
= H
4 + c c
H L + + 5 2 + c H + 2 + c L + 4λ 2 + 2 −6 H −4 H
cL 4 4 cL 4 cL 4 σH σL cL c3L
  
c3 c4H 1 1 c2H 1 1 cH 5 c2H 1
+9 H − 2 − 4λ + − c c
H L + − 2 + c H − − c L
c2L c3L σH2 σL2 c2L 2 2 cL 2 cL 2
5 1 cH 9 1 c2
− 3cH cL − c2H cL + c3H + cH c2L − 2 + cH − cL − 12 H
2 2 cL 2 2 cL
i.e.,
 2   
1 1 1 c2 cH c2H
4λ2 2 − σ2 − 2λ 2 c2H + 2 + 3cH + 2 H − c c
H L − 4 − 2 − c L
σH L σL c2L cL cL
 
1 2 c2H cH c2H
+ 2 cL + 2 − 3cL + 2 2 − cH cL − 4 + 5cH − 2
σH cL cL cL
c4H c2H 2 c4H c3H c3H c3H c4H c2H
=− − 5 − 9c H − + 6 + 4 − 9 + 2 + 12 + 4cH cL + 2c2H cL − c3H
c4L c2L c2L cL c3L c2L c3L cL
cH
− cH c2L + 2 − 5cH . (4.4)
cL
Summing up, the dispersion relation Dλ = 0 is equivalent to the system
(
eqn. (4.4)
Re(λ) ≤ cH,L
provided that conditions (4.1) and (4.3) are both satisfied.
We turn back to the equation (4.4) and we begin by observing that the term in square brackets
can be factorized as follows
(cL − cH ) 2
2 σ 2 c2 − σH cH c2L + 2σH2
cH cL − 2σH2 2 2
cH − σH 2
cL + 2σH 2
cL + 2σL 2
cH cL − 2σL cH
σH L L
2 3 2 2 2


+ σL cL − 3σL cL + 2σL cL . (4.5)
Clearly, if (4.5) is nonnegative, then (4.4) admits a solution with nonnegative part.
Now, we suppose now that (4.5) is negative and compute the discriminant of (4.4)
  
1 2 c2H cH c2H
∆ = 2 cH + 2 + 3cH + 2 2 − cH cL − 4 −2 − cL
σL cL cL cL
 2
1 c2 cH c2H
+ 2 c2L + 2 − 3cL + 2 H − c c
H L − 4 + 5c H − 2
σH c2L cL cL
 2  4 2 4 3
1 1 c cH 2 cH cH c3H c3H c4H c2H
−4 2 − 2 − H − 5 − 9c H − + 6 + 4 − 9 + 2 + 12
σH σL c4L c2L c2L cL c3L c2L c3L cL
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 19


cH
+ 4cH cL + 2c2H cL − c3H − cH c2L +2 − 5cH
cL

(cH − cL )2 1 2 2 2 4 3 2
=
c4L 4 (cL − 1) (8cH + 8cH cL − 16cH cL + cL − 4cL + 4cL )
σH
1
+ 4 (c2H c4L − 4c2H c3L + 12c2H c2L − 16c2H cL + 8c2H + 6cH c4L − 24cH c3L
σL
+ 32cH c2L − 16cH cL + c4L − 4c3L + 4c2L )

2 2 2 4 2 4 3 2
− 2 σ2 (c L − 1)(2c c
H L + c H c L − 4c H c L + c L − 4c L + 4c L .
) (4.6)
σH L
Clearly, if ∆ ≥ 0, then the equation (4.4) does not admit solutions with nonnegative real part.
On the other hand, if ∆ > 0, such an equation admits a real solution with nonnegative real part
if and only if and only if the term on the right-hand side of (4.4) is nonnegative. Such a term
can be factorized as follows:
cH (cL − 1)2 (cH − cL )2 (cH + c2L − 2cL )
−
c4L
Summing up, we have proved that the condition to have roots of the dispersion relation with
nonnegative real part are


 (cL − 2)(cH − cL ) ≥ 0,


(cL − 1)(2c2H + cH c2L − 4cH cL + c2L ) ≤ 0,

 (cL − cH )(−σH 2
cH c2L + 2σH 2
cH cL − 2σH 2 2 2
cH − σH 2
cL + 2σH 2
cL + 2σL 2
cH cL − 2σL cH

 2 3 2 2 2
+σL cL − 3σL cL + 2σL cL ) ≥ 0
or


 (cL − 2)(cH − cL ) ≥ 0,


 2 2 2
 (cL − 1)(2cH + cH cL − 4cH cL + cL )≤0,


 2
(cL − cH )(−σH cH c2L + 2σH 2
cH cL − 2σH 2 2 2
cH − σH 2
cL + 2σH 2
cL + 2σL 2
cH cL − 2σL cH
2 3 2 2 2

 +σ c
L L − 3σ c
L L + 2σL Lc )<0



 c (c − 1) 2
(c − c )2
(c + c 2
− 2cL )≤0,


H L H L H L

∆ > 0.
The proof is complete. 
Corollary 4.2. In the domain (A)  of financially
n 2 admissible 2parameters,oithe spectrum of the
c c
operator L consists of the halfline −∞, − min 8L (c2L − 1)2 , 8H (c2H − 1)2 .

Proof. By condition (3.1), it holds that cL < 1 < cH , so that (cL − 2)(cH − cL ) < 0. Hence, the
dispersion relation has no solutions and the statement follows from Theorem 3.4. 

5. Nonlinear stability analysis

In this section, we analyze the nonlinear Cauchy problem (2.20).
For further use, we introduce some notation. For every T ∈ [0, +∞) × {+∞}, by IT we
denote the interval [0, T ], if T ∈ R, and the halfline [0, +∞) otherwise. Moreover, I− and I+
denote, respectively, the intervals (−∞, η ∗ ] and [η ∗ , +∞). Finally, for every ω > 0 and every
♯ v1 (t, x)
pair v = (v1 , v2 ), with v1 : IT × I− → R and v2 : IT × I+ → R, we set v1,ω (t, x) = eωt
qL (x)
♯ ωt w1 (t, x) ♯ ♯ ♯
and v2,ω (t, x) = e and, then vω = (v1,ω , v2,ω ).
qH (x)
Let us introduce the following function spaces.
Definition 5.1. For every T ∈ (0, +∞) ∪ {+∞}, every ω ≥ 0 and every α ∈ (0, 1), the space
20 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

♯
• Xα,0 (T, ω) is the set of all pairs v = (v1 , v2 ) such that v1,ω ∈ Cbα,0 (IT × I− ; R) and
♯
v2,ω ∈ Cbα,0 (IT × I+ ; R). It is endowed with the norm
♯ ♯
kvkXα,0 (T,ω) = kv1,ω kC α,0 (IT ×I− ;R) + kv2,ω kC α,0 (IT ×I+ ;R) ;
b b

♯ α/2,α
• Xα/2,α (T, ω) is the set of all pairs v = (v1 , v2 ) such that v1,ω ∈ Cb (IT × I− ; R) and
♯ α/2,α
v2,ω ∈ Cb (IT × I+ ; R). It is endowed with the norm
♯ ♯
kvkXα/2,α(T,ω) = kv1,ω kC α/2,α (IT ×I− ;R) + kv2,ω kC α/2,α (IT ×I+ ;R) ;
b b

♯ 1+α/2,2+α
• X1+α/2,2+α (T, ω) is the set of all pairs v = (v1 , v2 ) such that v1,ω ∈ Cb (IT ×
♯ 1+α/2,2+α
I− ; R) and v2,ω ∈ Cb (IT × I+ ; R). It is endowed with the norm
♯ ♯
kvkX1+α/2,2+α (T,ω) = kv1,ω kC 1+α/2,2+α (IT ×I− ;R) + kv2,ω kC 1+α/2,2+α (IT ×I+ ;R) .
b b

Sometimes, we find it useful to denote by kvkX0,0 (T,ω) the sum kv1♯ kCb (IT ×I− ) + kv2♯ kCb (IT ×I+ ) .
To begin with, we observe thatthe function F is well defined in the subset  of X2 of pairs
 2
′ ∗ η∗ 2δ ∗ ∗
(w1 , w2 ) such that 1 − Ψ (w1 (η )) γe 1 − σ2 Ψ(w1 (η )) + Dy w1 (η− ) 6= 0 and
L

2
σL
′
Ψ′ (w1 (η ∗ ))Dyy w1 (η ∗ ) + δ − 12 σL
2
2
Ψ (w1 (η ∗ ))Dy w1 (η ∗ )
F (w) =   2 
′ ∗ η ∗ 2δ ∗ ∗
1 − Ψ (w1 (η )) γe 1 − σ2 Ψ(w1 (η )) + Dy w1 (η− )
L
  2 
∗ 2δ
× Dy w + γeη 1 − 2 Ψ(w) , (5.1)
σL
for such functions w. Since Ψ(0) = 0, it follows immediately that F is well defined in B(0, ρ) ⊂
D(L) for a sufficiently small ρ. In the following lemma we prove some properties of this function
and of function G, defined by
 
∗ η∗ 2δ 2δ 2 2
G(w) = R(w1 (η ))γe 1− 2 2 2 (σL − σH ), (5.2)
σL σH σL
for every w as above, where R is a smooth function (see (2.14)), which are crucial in the stability
analysis.
Lemma 5.2. There exists ρ0 > 0 such that for every ρ ∈ (0, ρ0 ], every T ∈ (0, +∞) ∪ {∞} and
ω ∈ (0, ∞), the following properties are satisfied.
(i) The function F maps the ball B(0, ρ) of X1+α/2,2+α (T, ω) into Xα/2,α (T, ω). Moreover,
there exists a positive constant C such that
kF (w2 ) − F (w1 )kXα/2,α (T,ω) ≤ Cρkw2 − w1 kX1+α/2,2+α (T,ω) (5.3)
for every w1 , w2 ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω).
(ii) The function G maps the ball B(0, ρ) of X1+α/2,2+α (T, ω) into C (1+α)/2 ([0, T ]). More-
over, there exists a positive constant C such that
keω· (G(w2 ) − G(w1 ))kC (1+α)/2 ([0,T ]) ≤ Cρkw2 − w1 kX1+α/2,2+α (T,ω) (5.4)
for every w1 , w2 ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω).
Proof. To prove the estimate in (i) and (ii), we need some preliminary results. More precisely,
we need to show that X1+α/2,2+α (T, ω) is continuosly embedded both into Xα/2,α (T, 0) and into
X(1+α)/2.0 (T, 0).
To begin with, we show that X1+α/2.2+α (T, ω) is continuosly embedded into Xβ,0 (T, 0) for
every β ∈ (0, 1). For this purpose, we recall that, if g ∈ C 1 ([0, T ]), then
|g(t2 ) − g(t1 )| ≤ kg ′ k∞ |t2 − t1 | ≤ kg ′ k∞ |t2 − t1 |β
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 21

for every t1 , t2 ∈ IT , with |t2 − t1 | ≤ 1, and

|g(t2 ) − g(t1 )| ≤ 2kgk∞ ≤ kgk∞ |t2 − t1 |β
for every t1 , t2 ∈ IT , with |t2 − t1 | ≥ 1. Hence,
[v]Xβ,0 (T,0) ≤2keω· vk∞ + kDt (e−ω (eω· v))k∞
≤2keω· vk∞ + kDt (eω· v)k∞ + ωkeω·vk∞
≤(2 + ω)kvkX1+α/2,2+α (T,ω) (5.5)
so that
kvkXβ,0 (T,0) ≤ (3 + ω)ρ. (5.6)
Next, we show that if w ∈ X1+α/2,2+α (ω.T ), then the spatial α-Hölder seminorm of w can be
estimated in a completely similar (and easier) way. It turns out that
♯ ♯ ♯
[w1,0 (t, ·)]Cbα (I− ) ≤2kw1,0 kCb (I− ) + kDy w1,ω kCb (I− )
1 ♯ Dy w1 (t, ·)
≤ (5 − cL )kw1,0 (t, ·)kCb (I− ) + (5.7)
2 qL Cb (I− )

and
♯ 1 ♯ Dy w(t, ·)
[w2,0 (t, ·)]Cbα (I+ ) ≤ (3 + cH )kw2,0 (t, ·)kCb (I+ ) + (5.8)
2 qH Cb (I+ )

for every t ∈ IT .
Combining estimates (5.6), (5.7) and (5.8), we deduce that w ∈ Xα/2,α (T, 0) and
kwkXα/2,α(T,0) ≤ c∗ kwkX1+α/2,2+α(T,ω)
for some positive constant c∗ , independent of w. which can be explicitly computed using the
quoted estimates.
Finally, we show that, if u ∈ X1+α/2,2+α (T, ω), then Dy u ∈ Xα/2,α (T, 0). For this purpose
we recall that there exists a positive constant c0 such that
1 1
kgkC 1 (I± ) ≤ c0 kgkC(I
2
±)
kgkC2 2 (I± ) , g ∈ C 2 (I± ),
so that we can estimate
Dy w1 (t2 , x) Dy w1 (t1 , x)
−
qL (x) qL (x)
♯ ♯ 1 ♯ ♯
≤|Dy w2,0 (t2 , x) − Dy w2,0 (t1 , x)| + (1 − cL )|w1,0 (t2 , x) − w1,0 (t1 , x)|
 2 
√ ♯
1
♯
1 1 ♯ α
≤ 2c0 kDt w1,0 kC2 b ([0,T ]×I− ) kw1,0 kC2 2 ([0,T ]×I− ) + (1 − cL )kDt w1,0 kCb ([0,T ]×I− ) |t − s| 2
b 2
for every x ∈ I− and every s, t ∈ IT with |t − s| ≤ 1. Moreover,
Dy w1 (t2 , x) Dy w1 (t1 , x) Dy w1 α
− ≤2 |t − s| 2
qL (x) qL (x) qL C([0,T ]×I− )

for every x ∈ I− and s, t ∈ IT , with |t − s| ≥ 1. A complete similar estimate can be proved when
x ∈ [η ∗ , +∞), with 1 − cL being replaced by cH − 1. Further, arguing as in (5.7) and (5.8) it can
be shown that
   
Dy w1 (t, ·) Dy w1 (t, ·)
+
qL C α (I− ) qH C α (I+ )
b b

1
≤ max{5 − cL , 3 + cH }kDy w(t, ·)kX0,0 (T,ω) + kDyy w(t, ·)kX0,0 (T,ω) .
2
22 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

for every t ∈ [0, T ]. Combining all these estimates, we conclude that there exists a positive
constant c∗∗ such that
kDy wkXα/2,α (T,0) ≤ c∗∗ kwkX1+α/2,2+α (T,ω) , w ∈ X1+α/2,2+α (T, ω). (5.9)
(i) Let ρ0 > 0 be such that
  2 
′ ∗ η∗ 2δ ∗ ∗ 1
1 − Ψ (w1 (t, η )) γe 1− 2 Ψ(w1 (t, η )) + Dy w1 (t, η− ) ≥
σL 2
for every w ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω) and every ρ ∈ (0, ρ0 ]. From now on ρ is arbitrarily
fixed in the interval (0, ρ0 ].
To begin with, we observe that, for every v ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω), it holds that
σL2
′
Ψ′ (w1 (t, η ∗ ))Dyy eωt w1 (t, η ∗ ) + δ − 21 σL
2
Ψ (w1 (t, η ∗ ))eωt Dy w1 (t, η ∗ )
eωt F (w(t, ·)) = 2   2 
′ ∗ η ∗ 2δ ∗ ∗
1 − Ψ (w1 (t, η )) γe 1 − σ2 Ψ(w1 (t, η )) + Dy w1 (t, η− )
L
  2 
∗ 2δ
× Dy w + γeη 1 − 2 Ψ(w)
σL
Fe1 (w(t, ·)) e
= F3 (w(t, ·))
Fe2 (w(t, ·))
for every t ∈ IT .
Using (5.5), (5.6) and observing that
Z 1
Ψ′ (w1 (·, η ∗ )) − Ψ′ (z1 (·, η ∗ )) = (w1 (·, η ∗ ) − z1 (·, η ∗ )) Ψ′′ (σw1 (·, η ∗ ) + (1 − σ)z1 (·, η ∗ ))dσ
0
we deduce that
kΨ(j) (w1 (·, η ∗ )) − Ψ(j) (z1 (·, η ∗ ))kC α/2 (IT )
b
 (j+1) 1 
≤ kΨ k∞,ρ + kΨ(j+2) k∞,ρ q∗ ([w]Xα/2,0 (T,0) + [z]Xα/2,0 (T,0) ) q∗ kw − zkXα/2,0 (T,0)
2
 
≤ kΨ(j+1) k∞,ρ + (2 + ω)ρkΨ(j+2) k∞,ρ q∗ (3 + ω)q∗ kw − zkX1+α/2,2+α (T,ω)
= : Cj q∗ kw − zkX1+α/2,2+α (T,ω) (5.10)

for every w, z ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω) and every j = 0, 1, where kΨ(k) k∞,ρ denotes the sup-
norm of the k-th derivative of the function ψ on the interval [0, ρq∗ ] and q∗ = max{qL (η ∗ ), qH (η ∗ )}.
Analogously,
kΨ(j) (w1 (·, η ∗ ))kC α/2 (IT ) ≤kΨ(j) k∞,ρ + kΨ(j+1) k∞,ρ q∗ (2 + ω)kwkX1+α/2,2+α(T,ω)
ej
≤kΨ(j) k∞,ρ + kΨ(j+1) k∞,ρ q∗ (2 + ω)ρ =: C (5.11)
for every w ∈ B(0, ρ) ⊂ X1+α/2,2+α (T, ω).
Straightforward computations based on (5.5), (5.6) and (5.9)-(5.11), reveal that
1e  2 
kFe1 (w)kC α/2 ([0,T ]) ≤ C 2
1 q∗ σL kDyy wkXα/2,0 (T,ω) + |2δ − σL |kDy wkXα/2,0 (ω.T )
2
1e  2 2

≤ C1 q∗ σL + |2δ − σL |c∗ ρ =: K1 ρ
2

kFe1 (w) − Fe1 (z)kC α/2 ([0,T ])

1  2 
≤ C1 q∗2 σL 2
kDyy wkXα/2,0 (T,ω) + |2δ − σL |kDy wkXα/2,0 (ω.T ) kw − zkX1+α/2,2+α (T,ω)
2
1e  2 2

+ C 1 q∗ σL kDyy w − Dyy zkXα/2,0 (T,ω) + |2δ − σL |kDy w − Dy zkXα/2,0 (T,ω)
2
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 23

 
1
≤ C1 q∗2 [σL
2 2
+ |2δ − σL |c∗∗ ]ρ + K1 kw − zkX1+α/2,2+α (T,ω)
2
= : K2 (ρ)kw − zkX1+α/2,2+α (T,ω) ;
  2 
1 e1 γeη∗ 1 − 2δ kΨk∞,ρ + q∗ kDy wkX (T,0)
≤2 + 4C 2
f2 (w)
F C α/2 ([0,T ]) σL 0,0

  
′ η∗ 2δ 2 e
+ 4kΨ k∞,ρ γe 1 − 2 C0 + q∗ [Dy w]Xα/2,0 (T,0)
σL
   
e η∗ 2δ 2 e
≤2 + 4C1 2γe 1 − 2 C0 + q∗ c∗∗ ρ =: K3 (ρ);
σL

1 1
−
f f
F2 (w) F2 (z) C α/2 ([0,T ])
  
∗ 2δ 2 e0 + C
e1 C0 ) + (C1 q∗ + C
e1 ρ)c∗∗ kw − zkX
≤q∗ (K3 (ρ))2 γeη 1 − 2 (C1 C 1+α/2,2+α (T,ω)
σL
= : K4 (ρ)kw − zkX1+α/2,2+α (T,ω) .
To estimate F3 (w), we observe that
Z 1
Ψ(w1 (·, x)) Ψ(z1 (·, x)) ♯ ♯
− =(w1,0 (·, x) − z1,0 (·, x)) Ψ′ (σw1 (·, x) + (1 − σ)z1 (·, x))dσ
qL (x) qL (x) 0
♯ ♯
= : (w1,0 (·, x) − z1,0 (·, x))J (·, x) (5.12)
for every x ∈ I− .
From this formula and arguing as in the proof of (5.10), it can be easily shown that
1 ′′ ♯ ♯
[J (·, x)]C α/2 (IT ) ≤ kΨ k∞,ρ (w1,0 (·, x)]C α/2 (IT ) + [z1,0 (·, x)]C α/2 (IT ) )q∗ (5.13)
b 2 b b

for every x ∈ (−∞, η ∗ ] and

1 ′′

[J (t, ·)]Cbα (I− ) ≤ kΨ k∞,ρ [w1 (t, ·)]C α (I− ) + [z1 (t, ·)]C α (I− ) , t ∈ IT .
2
Observing that
♯ ♯
[w1 (t, ·)]Cbα (I− ) ≤ [w1,0 (t, ·)]Cbα (I− ) q∗ + kw1,0 (t, ·)k∞ [qL ]Cbα (I− )
and
Z x2
1 1
|qL (x2 ) − qL (x1 )| ≤ (1 − cL ) e− 2 (cL −1)r dr
2 x1
 Z x2 1−α
1 α
(cL −1)r
− 2(1−α)
≤ (1 − cL )|x2 − x1 | e dr
2 x1
 α  α
1 − cL 1−α α 1 − cL
≤ (1 − α) |x2 − x1 | ≤ |x2 − x1 |α (5.14)
2 2
for every x1 , x2 ∈ (−∞, η ∗ ]. Hence,

1 ′′ ♯ ♯

[J (t, ·)]Cbα (I− ) ≤ kΨ k∞,ρ q∗ [w1,0 (t, ·)]Cbα (I− ) + [z1,0 (t, ·)]Cbα (I− )
2
 α 
1 − cL ♯ ♯

+ kw1,0 (t, ·)k∞ + kz1,0 (t, ·)k∞ (5.15)
2
From (5.12)-(5.15), we conclude that
Ψ(w1 ) Ψ(z1 )
−
qL qL Cb
α/2,α
(IT ×I− )
24 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

♯ ♯
≤kw1,0 − z1,0 kC α/2,α (IT ×I− ) kΨ′ k∞,ρ
b

1 ′′ ♯ ♯ ♯ ♯
+ kΨ k∞,ρ kw1,0 − z1,0 k∞ sup [w1,0 (·, x)]C α/2 (IT ) + sup [z1,0 (·, x)]C α/2 (IT )
2 x≤η ∗ b x≤η ∗ b

♯ ♯
+ sup [w1,0 (t, ·)]Cbα (I− ) + sup [z1,0 (t, ·)]Cbα (I− ) .
t∈IT t∈IT

Ψ(w1 ) Ψ(z1 )
In a completely similar way, we can estimate the function qH − qH .
Putting everything together and observing that
♯ ♯ ♯
[e−ω· w1,0 (·, x)]C α/2 (IT ) ≤[w1,0 (·, x)]C α/2 (IT ) + kw1,0 k∞ [e−ω· ]C α/2 (IT )
b b b
♯ ♯
≤[w1,0 (·, x)]C α/2 (IT ) + kw1,0 k∞ ω 1−α ,
b

where we have adapted (5.14) to estimate [e−ω· ]C α/2 (IT ) , and a similar estimate holds true with
w1 being replaced by z1 , we conclude that
kFf3 (w) − F f3 (z)kX
α/2,α (T,0)
  2  
η ∗ 2δ ′ ′′ α
≤ c∗ + γe 1 − 2 kΨ kρ.∞ c∗ + (2 + ω)ρq∗ kΨ k∞,ρ + 2ρcH,L kw − zkX1+α/2,2+α (T,ω)
σL
= : K5 (ρ)kw − zkX1+α/2,2+α (T,ω) ,

where cH,L = max 1−c L cH −1
2 , 2 .
Summing up, we have proved that
kF (w) − F (z)kXα/2,α (T,ω) ≤ K6 (ρ)ρkw − zkX1+α/2,2+α (T,ω) ,
 
where K6 (ρ) = K2 (ρ)K3 (ρ) + ρK1 (K3 (ρ) + K4 (ρ)) K5 (ρ) and (5.3) follows at once.
(ii) We observe observe that, since R(0) = R′ (0) = 0, we can write
Z 1
eω· [R(w1 (·, η ∗ )) − R(z1 (·, η ∗ ))] =eω· (w1 (·, η ∗ ) − z1 (·, η ∗ )) R′ (σw1 (·, η ∗ ) + (1 − σ)z1 (·, η ∗ ))dσ
0
Z 1
=eω· (w1 (·, η ∗ ) − z1 (·, η ∗ )) (σw1 (·, η ∗ ) + (1 − σ)z1 (·, η ∗ ))dσ
0
Z 1
× R′′ (τ σw1 (·, η ∗ ) + τ (1 − σ)z1 (·, η ∗ ))dτ.
0
(5.16)
Using the second equality in (5.16) to estimate the sup-norm of the function eω· [R(w1 (·, η ∗ )) −
R(z1 (·, η ∗ ))] and the first equality to estimate its (1 + α)-Hölder seminorm, we easily conclude
that
keω· [R(w1 (·, η ∗ )) − R(z1 (·, η ∗ ))]kC (1+α)/2 (IT )
b

1
≤ q∗2 kw − zkX(1+α)/2,0 (T,0) kR′′ k∞,ρ (kwkX(1+α)/2,0 (T,0) + kzkX(1+α)/2,0 (T,0) ).
2
Taking (5.6) into account, we from the previous estimate we deduce that
kG(w) − G(z)kC (1+α)/2 (IT )
b
 
2 η ∗ 2δ δ 2 2 ′′
≤q∗ c∗∗ γe 1− 2 2 σ 2 (σL − σH )kR k∞,ρ (2 + ω)ρkw − zkX(1+α)/2,0 (T,ω)
σL σH L

and (5.4) follows.

The proof is now complete. 

Based on these remarks, we are now in a position to prove the following theorem.
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 25

Theorem 5.3. Fix α ∈ [0, 1). Then,the nulln solution to problem (2.20)ois stable in the X2+α -
c2L 2 2
2 cH 2 2
norm. More precisely, for every ω0 ∈ 0, min 8 (cL − 1) , 8 (cH − 1) there exists a positive
constant ρ such that, for every initial datum w0 ∈ B(0, ρ) ⊂ X2+α , satisfying the compatibility
conditions
B(w0 ) = G(w0 ), B0 (Lw0 + F (w0 )) = 0,
the solution w to problem (2.20), with initial datum w0 , exists for all the positive times, belongs
to X1+α/2,2+α (+∞, ω0 ) and there exists a positive constant C such that
kw(t, ·)kX2+α ≤ Ce−ω0 t kw0 kX2+α , t > 0. (5.17)
Proof. Since it is rather long, we split the proof into several steps. We provide the proof in the
case when α ∈ (0, 1), since in the case α = 0 is similar and even simpler. Hence, throughout the
proof, α is arbitrarily fixed in (0, 1) and we use the notation introduced at the beginning of this
section.
Step 1. Here, we prove that for every T > 0, the Cauchy problem
 dw

 dt (t, ·) = Lw(t, ·) + f (t, ·), t ∈ [0, T ],

(5.18)

 (Bw)(t) = (0, g(t)), t ∈ [0, T ],

w(0, ·) = w0 ,
admits a unique solution w ∈ X1+α/2,2+α (T, 0), for every w0 ∈ X2α , every f = (f1 , f2 ) ∈
1+α
Xα/2,α (T, 0) and for every function g ∈ C 2 ([0, T ]; R), satisfying the compatibility conditions
Bw0 = g(0), B0 (Lw0 + f (0)) = 0.
Moreover, there exists a positive constant C such that


kwkX1+α/2,2+α (T,0) ≤ C kw0 kX2+α + kf kXα/2,α(T,0) + kgkC 1+α ([0,T ]) , (5.19)
where the constant C depends on only on the quantity Mk = supt∈[0,T +1] ktk Lk etL kL(X ) (k =
0, 1, 2), Mk,β = supt∈[0,T +1] ktk−β Lk etL kL(DL (β,∞),X ) (k = 0, 1, 2, 3) suitable choices of β.
For this purpose, we begin by introducing a lifting operator N , defined by N a = (aζ, 0) for
every a ∈ R, where ζ is a smooth function, which vanishes at x = η ∗ and satisfies the condition
ζ ′ (η ∗ ) = 1. A straightforward computation reveals that BN a = (0, a) for every a ∈ R.
The candidate to be the solution to the above Cauchy problem is the function w, defined by
Z t Z t
w(t, ·) = etL w0 + e(t−s)L f (s, ·) + LN g(s, ·))ds − L e(t−s)L (N g(s, ·))ds, (5.20)
0 0
for every t ∈ [0, T ].
To check all the above properties, we begin by introducing the function z : [0, T ] → X , defined
by
Z t
z(t) = e(t−s)L [N (g(s)) − N (g(0))]ds, t ∈ [0, T ].
0

Since the function t 7→ N (g(s)) − N (g(0)) belongs to C (1+α)/2 ([0, T ]; X2 ) which is contin-
uously embedded into C (1+α)/2 ([0, T ]; DL (1/2, ∞)), from [15, Theorem 4.3.16], it follows that
the function z belongs to C 1 ([0, T ]; X ) ∩ C 2 ([0, T ]; D(L)) and solves the equation z ′ (t) =
α
Lz(t) + N (g(t)) − N (g(0)) for every t ∈ [0, T ]. Moreover, Lz belongs to C 1+ 2 ([0, T ]; X ),
z ′ is bounded in [0, T ] with values in DL (1 + α/2, ∞) and there exists a positive constant C,
independent of w such that
kzkC 1 ([0,T ];X ) + kLzkC([0.T ];X ) + kz ′ kB([0,T ];DL (1+α/2,∞)) ≤ Ckgk 1+α (5.21)
C 2 ([0,T ])

By interpolation, it follows that z ′ is continuous with values in D(L). Writing,

Z t
z(t) − z(t0 ) = z ′ (s)ds, t, t0 ∈ [0, T ],
t0
26 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

and applying L to both sides of the previous equation, we conclude that

Z t
Lz(t) − Lz(t0 ) = Lz ′ (s)ds, t, t0 ∈ [0, T ].
t0
Hence, function Lz is differentiable in [0, T ] with values in X and (Lz)′ = Lz ′ = LLz +LN (g)−
LN (g(0)).
α
Let us now introduce the function w1 = −Lz + N (g(0)). Since (Lw1 )′ ∈ C 2 ([0, T ]; X ) ∩
B([0, T ]; DL (α/2, ∞)), the function (Lw1 )′ belongs to Xα/2,α (T, 0). As a byproduct, w1 belongs
to this space. Further, we observe that Lz = z ′ − N (g) + N (g(0)), so that Lz is bounded in
[0, T ] with values in X2+α , due to the characterization of the interpolation space DL (1 + α/2, ∞)
in Theorem 3.7. Moreover,
w1′ (t) = −(Lz)′ (t) = −LLz(t) − LN (g(t)) + LN (g(0)) = Lw1 (t) − LN (g(t))
for every t ∈ [0, T ].
To get the smoothness of the spatial derivatives of w1 = (w1 , w2 ) with respect to the time
variables, we write
Z t
w1 (t, x) − w1 (t0 , x) = Dt w1 (s, x)ds, t0 , t ∈ [0, T ] x ∈ (−∞, η ∗ ],
t0
which allows us to easily prove that
♯ ♯ ♯
kw1,0 (t, ·) − w1,0 (t0 , ·)kCbα (I− ) ≤C|t − t0 | sup kw1,0 (t, ·)kCbα (I− )
s∈[0,T ]

≤C|t − t0 |kDt w1 kXα/2,α (T,0)

for some positive constant C, independent of w1 . In the same way we can show that
♯ ♯
kw2,0 (t, ·) − w2,0 (t0 , ·)kCbα (I+ ) ≤ C|t − t0 |kDt w2 kXα/2,α(T,0) .
Since X2 belongs to the class J1−α/2 between Xα and X2+α , it follows that there exist positive
e and C1 , independent of w1 , such that
constants C
α
e 1 (t, ·) − w1 (t0 )k 2 kw1 (t, ·) − w1 (t0 )k 2 1= α
kw1 (t, ·) − w1 (t0 )kX2 ≤Ckw Xα X2+α
1− α α
≤C1 sup kw1 (t, ·)kX2+α
2
kDt w1 kXα/2,α (T,0) |t − t0 | 2
t∈[0,T ]

for every t, t0 ∈ [0, T ]. From estimate (5.21) and the previous arguments it follows that
kw1 kX1+α/2,2+α(T,0) ≤ Ckgk 1+α (5.22)
C 2 ([0,T ])
for some positive constant C, independent of w1 .
Summing up, we have proved that w1 belongs to X1+α/2,2+α (T, 0). Moreover, since z ′ (t) ∈
D(L) for every t ∈ [0, T ] and z ′ = Lz + N (g(t)) − N (g(0)) = −w1 + N (g(t)), it follows that
0 = B(−w1 (t)+N (g(t)) = −Bw1 (t)+N (g(t)) = −Bw1 (t)+g(t) for every t ∈ [0, T ]. Therefore,
w1 is a strict solution to the Cauchy problem
 dw
 1
 dt (t, ·) = Lw1 (t, ·) − LN (g(t)), t ∈ [0, T ],



 (Bw1 )(t) = (0, g(t)), t ∈ [0, T ],

w1 (0, ·) = 0.
Finally, we consider the Cauchy problem
 dw
 2
 dt (t, ·) = Lw2 (t, ·) + f + LN (g(t)),
 t ∈ [0, T ],
(5.23)

 (Bw2 )(t) = 0, t ∈ [0, T ],

w2 (0, ·) = w0 − N (g(0)),
Since u0 ∈ X2 and B(w0 −N (g(0))) = Bw0 −(0, g(0)) = 0, due to the compatibility conditions,
it follows that w0 − N (g(0)) belongs to D(L). Moreover, f + LN (g(t)) ∈ Xα/2,α (T, 0) =
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 27

B([0, T ]; Xα )∩C α/2 ([0, T ]; X ) and B0 (L(w0 −N (g(0)))+f (0)+LN (g(0))) = B0 (Lw0 +f (0)) =
0, again by the compatibility conditions. By classical results for abstract Cauchy problems
associated with sectorial operators, we conclude that problem (5.23) admits a unique classical
solution w2 which actually belongs to X1+α/2,2+α (T, 0) and satisfies the estimate
kw2 kX1+α/2,2+α(T,0) ≤ C2 (kw0 kX2+α + kf kXα/2,α (T,0) ). (5.24)
The function w = w1 + w2 is easily seen to be the (unique) strict solution to the Cauchy
problem (5.18) and, in view of estimates (5.22) and (5.24), (5.19) follows at once. To complete
this step, we observe that
Z t
w1 (t) = − L e(t−s)L (N (g(s)) − N (g(0)))ds + N (g(0))
0
Z t Z t
=−L e(t−s)L N (g(s))ds + L e(t−s)L N (g(0))ds + N (g(0))
0 0
Z t
=−L e(t−s)L N (g(s))ds + etL N (g(0))
0

for every t ∈ [0, T ]. Moreover,

Z t
tL
w2 (t) = e (u0 − N (g(0))) + e(t−s)L [f (s) + LN (g(s))]ds
0

for every t ∈ [0, T ]. so that the function w = w1 + w2 is given by (5.20).

Step 2. In this step, we consider the nonlinear problem
 dw

 (t, ·) = Lw(t, ·) + F (w(t, ·)), t ∈ [0, T ],

 dt
(Bw)(t) = (0, G(w(t, ·))), t ∈ [0, T ], (5.25)




w(0, ·) = w0 ,
where F (w) and G(w) are defined by (5.1) and (5.2).
By the results in Step 1, the solution to the previous problem is a fixed point of the operator
Γ : Y → Y defined by
Z t Z t
tL (t−s)L
(Γ(w))(t, ·) =e w0 + e F (w(s, ·)) + LN G(w(s, ·)))ds − L e(t−s)L N G(w(s, ·))ds
0 0

for every t ∈ [0, T ], where Y = B(0, ρ) ⊂ X1+α/2,2+α (ω0 , ∞) endowed with the norm of
X1+α/2,2+α (ω0 , ∞), where ρ > 0 needs to be properly fixed.
Note that
Z t
eω0 t (Γ(w))(t, ·) =et(L+ω0 ) w0 + e(t−s)(L+ω0 ) eω0 s F (w(s, ·)) + LN (eω0 s G(w(s, ·))))ds
0
Z t
−L e(t−s)(L+ω0 ) N (eω0 s G(w(s, ·)))ds
0

Using estimates (5.3) and (5.4) and (5.19), where the constant C therein appearing can be
made independent of T since it depends only on Mk,T (k = 0, 1, 2) and MK,β,+∞ (k = 0, 1, 2, 3),
which can be estimated from above by = supt≥0 ktk (L+ω0 )k et(L+ω0 ) kL(X ) and supt≥0 ktk−β (L+
ω0 )k et(L−ω) kL(DL (β,∞),X ) , respectively (these constants are finite since the semigroup (et(L+ω) )
is of negative type) it is immediate to check that Γ is a contraction from Y into itself, pro-
vided that kw0 kX2+α ≤ ρ0 and ρ0 , ρ are sufficiently small. Hence, for every w0 ∈ B(0, ρ0 ) ⊂
X2+α , problem (5.25) admits a unique solution which belongs to X1+α/2,2+α (ω0 , +∞), with
X1+α/2,2+α (ω0 , +∞)-norm which does not exceed ρ. Using a standard unique continuation ar-
gument, it is easy to infer that u is is the unique solution which belongs to X1+α/2,2+α (ω0 , +∞).
Estimate (5.17) follows at once. 
28 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

6. Convergence to the attenuated traveling wave

In this section, collecting all the previous results we go back to the original problem (1.2) and
(1.3) and state the main result of the paper. To enlighten the notation given a function ψ, we
write ψ ♯ both to denote the functions qψL and qψH . No confusion may arise since when we use the
ψ
notation ψ ♯ on a left-halfline it means that we are considering the function qL and when the use
the same notation on a right-halfline it means that we are considering the function qψH .
 n 2 o
c c2
Theorem 6.1. For every ω0 ∈ 0, min 8L (c2L − 1)2 , 8H (c2H − 1)2 there exists ρ0 > 0 such
that, for each φ0 such that the function φ♯0 −K belongs to Cb2+α ((−∞, η(0)]) and to Cb2+α ([η(0), +∞))
and kψ0♯ − KkC 2+α ((−∞,η(0)]) + kψ0♯ − KkC 2+α ([η(0),+∞)) ≤ ρ0 , the Cauchy problem (1.2), (1.3)
b b
admits a unique solution (φ, η), such that in the variable t and y = x + ct − η(t) + η ∗ , u belongs
to X1+α/2,2+α (ω0 , +∞). Moreover, there exists a positive constant C such that
|s(t) + ct − η ∗ | ≤ Ce−ω0 t |φ0 (η(0)) − K(η(0))|, t ≥ 0, (6.1)
and

kφ♯ (t, ·) − e−rt K ♯ kC 2+α ((−∞,s(t)]) ≤Ce−σ− t kφ♯0 − K ♯ kC 2+α ((−∞,s(0)]) + kφ♯0 − K ♯ kC 2+α ([s(0),+∞)
b b b

and

kφ♯ (t, ·) − e−rt K ♯ kC 2+α ([s(t),+∞)) ≤ Ce−σ+ t kφ♯0 − K ♯ kC 2+α ((−∞,s(0)]) + kφ♯0 − K ♯ kC 2+α ([s(0),+∞)
b b b

for every t ≥ 0, where σ− = − 1−C

2 c + ω0 + r and σ+ =
L 1−cH
2 c + ω0 + r are positive numbers.
Proof. Throughout the proof, by C we denote a positive constant, which can vary from line to
line but it is independent of t ≥ 0 and the functions that will appear.
To begin with, we observe that, by (2.7) and(2.13), the smoothness of Ψ and Theorem 5.3, it
follows that the function v = (v1 , v2 ), where
vj (t, ·) = Ψ(w(t, η ∗ ))K ′ + wj (t, ·), t > 0, j = 1, 2, (6.2)
belongs to X1+α/2,2+α (ω, +∞) and
 
K′ K′
kv(t, ·)kX2+α ≤ |Ψ(w1 (t, η ∗ ))| + + Ce−ω0 t kw0 kX2+α
qH Cb2+α ((−∞,η ∗ ]) qL Cb2+α ([η ∗ ,+∞))
(6.3)
for every t ≥ 0. Now, since
Z 1
Ψ(w1 (t, η ∗ )) = Ψ(w1 (t, η ∗ )) − Ψ(0) = w1 (t, η ∗ ) Ψ′ (σw1 (t, η ∗ ))dσ, t≥0
0
we deduce that |Ψ(w1 (t, η ∗ ))| ≤ Ce−ω0 t kw0 kX for every t ≥ 0. Hence,
kv(t, ·)kX2+α ≤ Ce−ω0 t kw0 kX2+α , t ≥ 0.
To get rid of the norm of w0 from the right-hand side of the previous estimate, we denote by
w0 = (w0,1 , w0,2 ) the value of w at t = 0 and, similarly, by v0 = (v0,1 , v0,2 ) we denote the value
of v at t = 0. From formula (6.2) we can infer that v0 = w0 + Ψ(w0,1 (η ∗ ))K ′ , so that
 
∗ 2δ
v0,1 (η ∗ ) = w0,1 (η ∗ ) + Ψ(w0,1 (η ∗ ))K ′ (η ∗ ) = w0,1 (η ∗ ) + γeη 1 − 2 Ψ(w0,1 (η ∗ )).
σL
∗
 
Since 1 − σ2δ2 = 1 − cL > 0 by (3.1), the function x 7→ Λ(x) := x + γeη 1 − σ2δ2 Ψ(x) is
L L
strictly increasing in a neighborhood of the origin. Hence, Λ is invertible in a neighborhood of
the origin and we can write w0,1 (η ∗ ) = Λ−1 (v0,1 (η ∗ )). We have so proved that
w0 = v0 − Ψ(Λ−1 (v0,1 (η ∗ )))K ′
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 29

The functions Λ−1 and Ψ vanish at zero and are smooth in a neighborhood of the origin. From
this remark, we obtain that kw0 kX2+α ≤ Ckv0 kX2+α , so that, from (6.3), we deduce that
kv(t, ·)kX2+α ≤ Ce−ω0 t kv0 kX2+α , t ≥ 0. (6.4)
Next, we go back to function u, by means of formula (2.3), which shows that u(t, ξ) = v1 (t, ξ −
η(t) + η ∗ ) + K(ξ) if t ≥ 0 and ξ < η(t) and u(t, ξ) = v2 (t, ξ − η(t) + η ∗ ) + K(ξ) if t ≥ 0 and
ξ > η(t). Moreover, η(t) − η ∗ = f (t) = Ψ(w1 (t, η ∗ )) for every t ≥ 0, again due to formula (2.13).
1+α/2
It turns out that η ∈ Cb ([0, +∞)) and
|η(t) − η ∗ | ≤ C4 e−ω0 t |v0,1 (η ∗ )| = Ce−ω0 t |u0 (η(0)) − K(η ∗ )|, t ≥ 0. (6.5)
We have so proved estimate (6.1). Moreover, since qL (ξ) = qL (ξ − η(t) + η ∗ )qL (η ∗ − η(t)) and a
similar formula holds true with qL being replaced by qH , we can infer that
u(t, ξ) K(ξ) v1 (t, ξ − η(t) + η ∗ ) v1 (t, ξ − η(t) + η ∗ )
− = = qL (η(t) − η ∗ ),
qL (ξ) qL (ξ) qL (ξ) qH (ξ − η(t) + η ∗ )
for every t ≥ 0 and ξ ≤ η(t), and
u(t, ξ) K(ξ) v1 (t, ξ − η(t) + η ∗ )
− = qH (η(t) − η ∗ ),
qH (ξ) qH (ξ) qH (ξ − η(t) + η ∗ )
u u
functions qL and qH are smooth in space and time and

ku♯ (t, ·) − K ♯ kC 2+α ((−∞,η(t)]) ≤ kv1♯ (t, ·)kC 2+α ((−∞,η∗ ]) qL (η(t) − η ∗ ) ≤ Ce−ω0 t kv0 kX2+α
b b

ku♯ (t, ·) − K ♯ kC 2+α ([η(t),+∞))) ≤ kv2♯ (t, ·)kC 2+α ([η∗ ,+∞)) qH (η(t) − η ∗ ) ≤ Ce−ω0 t kv0 kX2+α
b b

for every t ≥ 0, where we took into account (6.4) and (6.5), which shows that the functions
qL (η − η ∗ ) and qH (η − η ∗ ) are bounded in [0, +∞). Since


kv0 kX2+α ≤ C ku♯0 − K ♯ kC 2+α ((−∞,η∗ ]) + ku♯0 − K ♯ kC 2+α ([η∗ ,+∞) ,
b b

where u0 = u(0, ·), from the previous two estimate we can infer that


ku♯ (t, ·) − K ♯ kC 2+α ((−∞,η(t)]) ≤ Ce−ω0 t ku♯0 − K ♯ kC 2+α ((−∞,η∗ ]) + ku♯0 − K ♯ kC 2+α ([η∗ ,+∞)
b b b


ku♯ (t, ·) − K ♯ kC 2+α ([η(t),+∞))) ≤ Ce−ω0 t ku♯0 − K ♯ kC 2+α ((−∞,η∗ ]) + ku♯0 − K ♯ kC 2+α ([η∗ ,+∞) .
b b b

Finally, we go back to problem (1.2), (1.3). Using condition (1.4), it is immediate to check
that s(t) = η(t)−ct fot every t ≥ 0 and φ(t, x) = v(t, x+ct) for every t ≥ 0 and x ∈ R. Therefore,
 
φ(t, x) K(x) u(t, x + ct) K(x + ct) u(t, x + ct) K(t, x + ct)
− = − = − qL (ct)
qL (x) qL (x) qL (x) qL (x + xt) qL (t, x + ct) qL (t, x + ct)
for every t ≥ 0 and x ≤ s(t) and
 
φ(t, x) K(x) u(t, x + ct) K(x + ct) u(t, x + ct) K(t, x + ct)
− = − = − qH (ct)
qH (x) qH (x) qH (x) qH (x + xt) qH (t, x + ct) qH (t, x + ct)
for every t ≥ 0 and x ≤ s(t). It thus follows that
kφ♯ (t, ·) − e−rt K ♯ kC 2+α ((−∞,s(t)]) ≤ku♯ (t, ·) − K ♯ kC 2+α ((−∞,η(t)] qL (ct)
b b

≤Ce −σ− t
kφ♯0 − K ♯ kC 2+α ((−∞,s(0)]) + kφ♯0 − K ♯ kC 2+α ([s(0),+∞)
b b

and
kφ♯ (t, ·) − e−rt K ♯ kC 2+α ([s(t),+∞)) ≤ku♯ (t, ·) − K ♯ kC 2+α ([s(t),+∞)) qL (ct)
b b

≤Ce−σ+ t kφ♯0 − K ♯ kC 2+α ((−∞,s(0)]) + kφ♯0 − K ♯ kC 2+α ([s(0),+∞)
b b

for every t ≥ 0, where σ− = − 1−C

2 c + ω0 + r and σ+ =
L 1−cH
2 c + ω0 + r.
30 C.-M. BRAUNER, Y. DONG, J. LIANG, AND L. LORENZI

Since c = r − δ we can write

1 − cL 1 − CL
σ− = r + ω0 + δ,
2 2

1 − cH 3 − cH cH − 1
σ+ = c + ω0 + r = r+ω+ δ
2 2 2
and the right-hand sides of the previous formulas are positive since we are assuming that cL ∈
(0, 1) and cH ∈ (1, 3] (see (3.1)). The proof is complete. 

acknowlegments
C.-M. B. greatly acknowledges the School of Mathematical Sciences of Tongji University for the
warm hospitality during his visiting position from 2019 to 2021. He also expresses his gratitude
to the Department of Mathematics and Computer Sciences of the University of Parma. C.-M.
B. and J. L. acknowledge partial support from National Natural Science Foundation of China
(No. 12071349). Y. D. acknowledges partial support from National Natural Science Foundation
of China (No. 12071333 & No. 12101458). L. Lorenzi is member of G.N.A.M.P.A. of the Italian
Istituto Nazionale di Alta Matematica (INdAM).

References
[1] C.J. Ammon, A.A. Velasco, T. Lay and T.C. Wallace, Foundations of Modern Global Seismology, Second
edition, Elsevier, 2020.
[2] D. Addona, C.-M. Brauner, L. Lorenzi and W. Zhang, Instabilities in a combustion model with two free
interfaces, J. Differential Equations 268 (2020), 396–4016.
[3] C.-M. Brauner, J. Hulshof, A. Lunardi, A general approach to stability in free boundary problems, J. Differ-
ential Equations 164 (2000), 16–48.
[4] C.-M. Brauner, L. Lorenzi, Local existence in free interface problems with underlying second-order Stefan
condition, Rev. Roumaine Math. Pures Appl. 23 (2018), 339–359.
[5] C.-M. Brauner, L. Lorenzi, Instability of free interfaces in premixed flame propagation, Discr. Cont. Dyn.
Syst. Series S 14 (2021), 575–596.
[6] C.-M. Brauner, L. Lorenzi, M. Zhang, Stability analysis and Hopf bifurcation for large Lewis number in a
combustion model with free interface. Ann. Inst. H. Poincaré Anal. Non Linéaire 37 (2020), 581–604
(2020)
[7] X.F. Chen, B. Hu, J. Liang, H.M. Yin, The free boundary problem for measuring credit rating migration
risks, Scientia Sinica Mathematica, 54, (2024), 1–28
[8] Y. Dong, J. Liang, C.-M. Brauner, Double free boundary problem for defaultable corporate bond with credit
rating migration risks and their asymptotic behaviors, J. Differential Equations 372 (2023), 505–535.
[9] B. Hu, J. Liang, Y. Wu, A Free Boundary Problem for Corporate Bond with Credit Rating Migration, J.
Math. Anal. Appl. 428 (2015) 896–909.
[10] R.A. Jarrow, D. Lando, S. M.Turnbull, A Markov model for the term structure of credit risk spreads, Review
of Financial studies 10 (1997), 481–523.
[11] R.A. Jarrow, S.M. Turnbull, Pricing Derivatives on Financial Securities Subject to Credit Risk, Journal of
Finance 50 (1995), 53–86.
[12] J. Liang, Y. Wu, B. Hu, Asymptotic traveling wave solution for a credit rating migration problem, J. Differ-
ential Equations 261 (2016), 1017–1045.
[13] J. Liang, Z.K. Zeng, Pricing on defaultable and callable bonds with credit rating migration risks under
structure framework, Applied Mathematics A Journal of Chinese Universities (Ser.A), 2015, 61-70.
[14] L. Lorenzi, A. Rhandi, Semigroups of bounded operators and second-order elliptic and parabolic partial
differential equations. Monogr. Res. Notes Math. CRC Press, Boca Raton, FL, 2021.
[15] A. Lunardi, Analytic Semigroups and Optimal Regularity in Parabolic Problems, Birkhäuser, Basel, 1996.
[16] A.A. Markov, Rasprostranenie zakona bol’shih chisel na velichiny, zavisyaschie drug ot druga, Izvestiya
Fiziko-matematicheskogo obschestva pri Kazanskom universitete, 2-ya seriya 15 (1906), 135–156.
[17] R.C. Merton, On the Pricing of Corporate Debt: The Risk Structure of Interest Rates, Journal of Finance
29 (1974), 449–470.
[18] D.H. Sattinger, On the stability of waves of nonlinear parabolic systems, Adv. Math. 22 (1976), 312–355.
 STABILITY IN A CREDIT RATING MIGRATION PROBLEM 31

Institut de Mathématiques de Bordeaux, Université de Bordeaux, 33405 Talence, France

Email address: [email protected]

School of Mathematical Sciences, Tongji University, Shanghai 200092, China

Email address: [email protected]

School of Mathematical Sciences, Tongji University, Shanghai 200092, China

Email address: liang [email protected]

Plesso di Matematica, Dipartimento di Scienze Matematiche, Fisiche e Informatiche, Università di

Parma, Parco Area delle Scienze 53/A, I-43124 Parma, Italy
Email address: [email protected]