Singularly perturbed differential equations arise in mathematical models in which processes evolving on widely separated time or spatial scales interact. Their defining feature is the presence of a small positive parameter multiplying the highest-order derivative. As this parameter approaches zero, the order of the differential equation is reduced, and the limiting problem can no longer accommodate all the conditions imposed on the original equation. This loss of order generally produces thin transition regions, known as boundary or initial layers, where the solution or its derivatives vary rapidly.

These problems occur in fluid mechanics, reaction kinetics, control theory, electrical circuits, transport phenomena, and other areas involving multiple timescales. From the analytical point of view, the principal difficulty is that convergence to the reduced solution is usually nonuniform near the layer. Although the solution itself may converge uniformly, its derivatives may retain an order-one discrepancy within a region whose thickness tends to zero with the perturbation parameter.

For linear ordinary differential equations, the qualitative convergence theory is well established [1]-[3]. Nevertheless, qualitative convergence alone does not determine the rate of approximation, the dependence of the estimates on the data, or the precise contribution of the rapidly decaying characteristic mode. These questions are essential for rigorous asymptotic analysis and for the construction of numerical methods whose accuracy remains uniform with respect to the perturbation parameter.

The purpose of this paper is to derive explicit and parameter-uniform error estimates for a second-order singularly perturbed Cauchy problem with constant coefficients. Our approach combines an exact Green’s function representation with a spectral decomposition of the characteristic roots into slow and fast modes. This makes it possible to identify the initial-layer contribution explicitly and to determine the compatibility condition under which that contribution disappears.

The main contributions of the paper are as follows:

(i) an explicit and rigorous O ( ε ) uniform estimate for the difference between the perturbed and reduced solutions;

(ii) a pointwise decomposition of the derivative error into an exponentially localized initial-layer term and an O ( ε ) remainder;

(iii) an exact compatibility condition ensuring uniform C 1 convergence;

(iv) numerical experiments comparing exact errors with the explicit theoretical bounds;

(v) a closed-form example demonstrating the sharpness of the estimates.

The remainder of the paper is organized as follows. Section 2 introduces the singularly perturbed problem and its reduced counterpart. Section 3 establishes the functional setting and regularity of the reduced solution. Section 4 analyzes the characteristic roots and the separation of timescales, while Section 5 derives the error equation and its Green’s-function representation. Sections 6 and 7 establish the uniform solution and derivative estimates. Section 8 states precise prototypes for extensions, Section 9 presents numerical experiments, and Section 10 gives an explicit example confirming sharpness. Conclusions are collected in Section 11.

Let C k ( [ 0 , b ] ) denote the Banach space of k -times continuously differentiable functions with the sup-norm ‖   ⋅   ‖ ∞ . Constants are defined locally in each result and are independent of ε ∈ ( 0 , ε adm ] .

Proposition3.1(Reduced Solution Regularity)Under f ∈ C 2 ( [ 0 , b ] ) ,the reduced problem(2.5)admits a unique solution y 0 ∈ C 3 ( [ 0 , b ] ) given explicitly by

Moreover, y 0 ( k ) = ( f ( k − 1 ) − a 0 y 0 ( k − 1 ) ) / a 1 for k = 1 , 2 , 3 .

Proof. Equation (2.5) is linear first-order with constant coefficient a 0 / a 1 . Multiplying by the integrating factor e ( a 0 / a 1 ) x ,

and integrating from 0 to x , using y 0 ( 0 ) = A 0 ,

which is (3.1). Since f ∈ C 2 , differentiating (2.5) once gives a 1 y ″ 0 + a 0 y ′ 0 = f ′ , so y ″ 0 = ( f ′ − a 0 y ′ 0 ) / a 1 ∈ C 1 ( [ 0 , b ] ) , i.e. y 0 ∈ C 3 ( [ 0 , b ] ) ; differentiating once more gives y ‴ 0 = ( f ″ − a 0 y ″ 0 ) / a 1 ∈ C ( [ 0 , b ] ) , confirming the stated recurrence for k = 1 , 2 , 3 . □

Theorem3.2(Well-Posednessof the FullProblem)Forevery fixed ε > 0 ,the problem(2.1)-(2.2)possesses a unique solution y ε ∈ C 2 ( [ 0 , b ] ) .

Proof. Writing Y = ( y ε , y ′ ε ) T , equation (2.1) is equivalent to the linear first-order system

The matrix is constant and f / ( ε a 2 ) ∈ C ( [ 0 , b ] ) , so the right-hand side is globally Lipschitz in Y and continuous in x . Standard existence and uniqueness theory for linear systems therefore gives a unique solution on [ 0 , b ] , with Y ∈ C 1 ( [ 0 , b ] ; ℝ 2 ) and hence y ε ∈ C 2 ( [ 0 , b ] ) . □

The characteristic polynomial of G ε = ε a 2 D 2 + a 1 D + a 0 is

with roots

Since ε ≤ ε adm ≤ a 1 2 / ( 8 a 0 a 2 ) , the discriminant is positive. Thus both roots are real, negative (by Vieta: r ε + + r ε − = − a 1 / ( ε a 2 ) < 0 and r ε + r ε − = a 0 / ( ε a 2 ) > 0 ), and distinct.

Proposition4.1(Rigorous Root Bounds and Asymptotic Expansion)Let κ 1 : = a 0 / a 1 and κ 2 : = a 1 / ( 2 a 2 ) . Then,for every ε ∈ ( 0 , ε adm ] ,

Moreover, as ε → 0 + , the roots admit the uniformly valid asymptotic expansions

Proof.Step1(an exact closed form for r ε + ,and the bound(4.2)). Rationalize the numerator of r ε + in (4.1) by multiplying and dividing by ( a 1 + a 1 2 − 4 ε a 2 a 0 ) :

which simplifies, after cancelling 2 ε a 2 ≠ 0 , to the exact, ε -uniform identity

For ε ∈ ( 0 , ε adm ] the quantity σ ε : = a 1 2 − 4 ε a 2 a 0 satisfies a 1 / 2 ≤ σ ε < a 1 , hence

which is exactly the bound on r ε + in (4.2), with no asymptotic approximation involved.

Step2(bound on r ε − ). By Vieta’s formula r ε + + r ε − = − a 1 / ( ε a 2 ) , so

using the lower bound r ε + ≥ − 2 a 0 / a 1 from Step 1 (so − r ε + ≤ 2 a 0 / a 1 ). The admissible threshold (2.4) implies

For such ε ,

which is the second bound in (4.2).

Step3(asymptotic expansions(4.3)-(4.4)). Write u : = 4 ε a 2 a 0 / a 1 2 → 0 and use the binomial (Taylor) expansion 1 − u = 1 − 1 2 u − 1 8 u 2 + O ( u 3 ) for | u | < 1 :

Substituting into r ε + = ( − a 1 + σ ε ) / ( 2 ε a 2 ) :

which is (4.3). For r ε − = ( − a 1 − σ ε ) / ( 2 ε a 2 ) ,

which is (4.4). (Alternatively, (4.4) follows instantly from (4.3) and Vieta’s identity r ε − = − a 1 / ( ε a 2 ) − r ε + , without a second Taylor expansion.) □

Remark4.2Only the exact bounds(4.2),not the asymptotic expansions(4.3)-(4.4),are needed to prove the error estimates of Sections5-6; the expansions are retained because they give the leading-order constants in the worked example of Section1.

Define the error z ε : = y ε − y 0 . Subtracting (2.5) from (2.1) and using a 1 y ′ 0 + a 0 y 0 = f ,

Therefore the exact error equation is

The Green’s function (impulse-response fundamental solution) of ℒ ε = ε a 2 D 2 + a 1 D + a 0 , i.e. the solution of ℒ ε G ε = 0 with G ε ( 0 ) = 0 , G ′ ε ( 0 ) = 1 / ( ε a 2 ) , is obtained from the two exponential modes e r ε + t , e r ε − t :

Indeed G ε ( 0 ) = 0 and G ′ ε ( 0 ) = ( r ε + − r ε − ) / ( ε a 2 ( r ε + − r ε − ) ) = 1 / ( ε a 2 ) , as required.

Lemma5.1(Uniform Bounds for the Green’s Function and Its Derivative)For every ε ∈ ( 0 , ε adm ] and t ∈ [ 0 , b ] ,estimates(5.3)-(5.5)hold with

These constants depend only on a 0 , a 1 , a 2 and are independent of ε , f , and the initial data.

Proof.Step1(lower bound for the denominator). By (4.1), ε a 2 ( r ε + − r ε − ) = a 1 2 − 4 ε a 2 a 0 = σ ε . Definition (2.4) gives 4 ε a 2 a 0 ≤ a 1 2 / 2 , hence σ ε ≥ a 1 / 2 > a 1 / 2 . Equivalently,

Thus both (4.2) and σ ε ≥ a 1 / 2 hold throughout ( 0 , ε adm ] ; no additional parameter threshold is introduced.

Step2(bound(5.3)). Since r ε − < r ε + < 0 by Proposition 4.1 and t ≥ 0 , both e r ε + t and e r ε − t lie in ( 0 , 1 ] , so

Hence, using ε a 2 ( r ε + − r ε − ) = σ ε ≥ a 1 / 2 ,

This proves (5.3) with the explicit constant C G = 2 / a 1 .

Step3(bound(5.4)). Differentiating (5.2),

Term(I). By (4.2), | r ε + | ≤ 2 a 0 / a 1 and σ ε ≥ a 1 / 2 , so

| r ε + σ ε | ≤ 2 a 0 / a 1 a 1 / 2 = 4 a 0 a 1 2 ,   e r ε + t ≤ e − κ 1 t ( κ 1 = a 0 / a 1 , by (4.2),

hence | ( I ) | ≤ 4 a 0 a 1 2 e − κ 1 t .

Term(II). Write | r ε − | / σ ε . From Vieta, r ε − = − a 1 / ( ε a 2 ) − r ε + , so | r ε − | ≤ a 1 / ( ε a 2 ) + 2 a 0 / a 1 ; since σ ε ≥ a 1 / 2 ,

for C 2 : = 2 / a 2 + 4 a 0 / a 1 2 , because ε ≤ ε adm ≤ 1 . By (0.8), e r ε − t ≤ e − κ 2 t / ε with κ 2 = a 1 / ( 2 a 2 ) . Hence | ( II ) | ≤ C 2 ε e − κ 2 t / ε .

Combining (I) and (II) with κ : = min { κ 1 , κ 2 } and C ′ G : = max { 4 a 0 / a 1 2 , C 2 } gives

which is (5.4).

Step4(integral bounds(5.5)). Directly, ∫ 0 b | G ε ( t ) | d t ≤ C G ∫ 0 b d t = C G b . For the second,

using ∫ 0 b e − κ t / ε d t = ε ( 1 − e − κ b / ε ) / κ ≤ ε / κ and bounding both terms by C ′ G / κ .

This is (5.5) (with the constant 2 C ′ G / κ in place of the schematic C ′ G ( 1 / κ + 1 / κ ) ). □

By the variation-of-parameters formula for the linear operator ℒ ε , the unique solution of (5.1) is

Remark5.2(Derivation of(5.6))Write z ε = z h + z p . The homogeneous part z h solves ℒ ε z h = 0 with z h ( 0 ) = 0 , z ′ h ( 0 ) = Δ ;since G ε itself already satisfies G ε ( 0 ) = 0 , G ′ ε ( 0 ) = 1 / ( ε a 2 ) ,the multiple z h ( x ) = ε a 2 Δ G ε ( x ) satisfies z h ( 0 ) = 0 and z ′ h ( 0 ) = ε a 2 Δ ⋅ 1 ε a 2 = Δ ,as required. Theparticular part z p solves ℒ ε z p = − ε a 2 y ″ 0 with zero initialdata;by Duhamel’s principle for the impulse response G ε ,

because the forcing term of (5.1) is exactly − ε a 2 y ″ 0 and G ε is, by definition, the response of ℒ ε to a unit impulse. Summing z h + z p gives (5.6).

Theorem6.1(Optimal O ( ε ) Uniform Error Bound)Under(2.3),for every ε ∈ ( 0 , ε adm ] ,

with C G = 2 / a 1 as in Lemma 5.1. In particular y ε → y 0 uniformly on [ 0 , b ] as ε → 0 + .

Proof. Fix x ∈ [ 0 , b ] . From the exact representation (5.6) and the triangle inequality,

By (5.3), | G ε ( x ) | ≤ C G for the first term. For the integral term, bound | y ″ 0 ( s ) | ≤ ‖ y ″ 0 ‖ ∞ (constant, taken out of the integral) and | G ε ( x − s ) | ≤ C G pointwise for s ∈ [ 0 , x ] , so

since x ≤ b . Adding the two bounds,

Since the right-hand side does not depend on x , taking the supremum over x ∈ [ 0 , b ] gives (6.1) with C = a 2 C G max { 1 , b } . □

Remark6.2Theorem6.1strengthens the qualitative convergence statement discussed in Section2.2by providing a parameter-uniform O ( ε ) bound valid for every ε ∈ ( 0 , ε adm ] .

Differentiating (5.6) in x (the boundary term at s = x inside the integral is handled by Leibniz’s rule, and vanishes because G ε ( 0 ) = 0 ):

Theorem7.1(Derivative Estimate with an Initial-Layer Term)For all x ∈ [ 0 , b ] and ε ∈ ( 0 , ε adm ] ,estimate(7.1)holds with

Thus C and κ depend only on the fixed coefficients and b , and are independent of ε , f , and the initial data.

Proof. We treat the two terms of z ′ ε ( x ) separately.

Step1: the homogeneous(initial-condition)term ε a 2 Δ G ′ ε ( x ) . From (5.2),

For the slow component, r ε + − r ε − = σ ε / ( ε a 2 ) ≥ a 1 / ( 2 ε a 2 ) and | r ε + | ≤ 2 a 0 / a 1 . Therefore

and | e r ε + x | ≤ 1 . For the fast component, the exact formulae for the roots give

where we used σ ε ≥ a 1 / 2 . Since e r ε − x ≤ e − κ 2 x / ε , the two components satisfy

for every x ∈ [ 0 , b ] and ε ∈ ( 0 , ε adm ] .

Step2: the convolution(particular)term. Integrate by parts in

by setting u = y ″ 0 ( s ) , d v = G ′ ε ( x − s ) d s , so d u = y ‴ 0 ( s ) d s and, since d d s [ − G ε ( x − s ) ] = G ′ ε ( x − s ) , we take v = − G ε ( x − s ) :

At s = x : G ε ( 0 ) = 0 , so this boundary contribution vanishes. At s = 0 : the term is − ( − G ε ( x ) y ″ 0 ( 0 ) ) = G ε ( x ) y ″ 0 ( 0 ) (mind the overall sign from evaluating the bracket at the lower limit with a minus sign in front). Hence

Using | G ε ( t ) | ≤ C G from (5.3) for both the boundary term and inside the integral,

Multiplying by ε a 2 ,