Discrete Fractional Gronwall Inequality Overview

• Discrete fractional Gronwall inequality is a family of comparison inequalities for grid functions using fractional kernels and Mittag–Leffler functions.
• It facilitates stability and error analysis in numerical methods for fractional differential equations, including L1 and convolution quadrature schemes.
• It employs auxiliary inverse kernels and resolvents to transform nonlocal recursions into explicit bounds under specific step-size conditions.

Discrete fractional Gronwall inequality denotes a family of comparison inequalities for nonnegative sequences or grid functions governed by nonlocal discrete operators that approximate fractional derivatives or fractional sums. In the same way that the classical discrete Gronwall inequality converts a recursive inequality into an exponential bound, the fractional version converts a discrete convolution inequality into a bound involving fractional kernels and, typically, the Mittag–Leffler function $E_\alpha$. The subject now spans L1 approximations of the Caputo derivative, convolution quadratures for Riemann–Liouville and Caputo operators, nonuniform-mesh formulations with complementary kernels, completely positive discretizations, and more abstract settings such as time scales, discrete Riemann–Liouville calculus, and $q$-fractional calculus (Li et al., 2016, Liao et al., 2018, Feng et al., 2024).

1. Continuous prototype and the fractional comparison principle

The modern discrete theory is closely tied to continuous fractional Gronwall inequalities. A representative continuous result is given for the Caputo–Katugampola framework, where the kernel

$\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$

is the kernel of the Katugampola fractional integral. If $u,v$ are nonnegative, $g$ is nonnegative and nondecreasing, and

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$

then one obtains a series bound and, when $v$ is nondecreasing, the closed form

$$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$

The proof proceeds by introducing a fractional Volterra operator $V$, iterating $u\le v+Vu$, estimating $q$0 by Beta- and Gamma-function identities, and proving $q$1. The same paper explicitly states that this continuous inequality provides the template for a discrete fractional Gronwall inequality by replacing the fractional integral operator with a discrete convolution operator on a grid (Almeida, 2017).

This continuous origin explains two enduring features of the discrete theory. First, the growth factor is usually Mittag–Leffler rather than exponential. Second, the decisive object is not merely a time-step recurrence, but a memory kernel induced by a fractional operator. In numerical analysis, this continuous-to-discrete passage is what turns an error equation for a fractional differential equation into a discrete comparison inequality.

2. Canonical discrete forms on uniform grids

A foundational uniform-grid result is the discrete fractional Gronwall inequality for the L1 approximation to the Caputo derivative. On the uniform partition $q$2, the L1 method uses weights

$q$3

and the main inequality states that if nonnegative sequences $q$4 and $q$5 satisfy

$q$6

then there exists a positive constant $q$7 such that, when $q$8,

$q$9

The proof introduces an auxiliary sequence $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$0, proves identities such as

$\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$1

and then rewrites the problem as a lower-triangular matrix inequality whose powers are controlled by Mittag–Leffler growth. This result was presented as the missing “fundamental Gronwall type inequality” for nonlinear L1 analyses (Li et al., 2016).

A second uniform-grid strand concerns backward Euler convolution quadrature for the Riemann–Liouville derivative. There the discrete operator is

$\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$2

with Grünwald–Letnikov weights generated by $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$3. If nonnegative sequences $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$4, $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$5 satisfy

$\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$6

then for $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$7,

$\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$8

Here the analysis relies on a second auxiliary sequence $\tau^{p-1}(t^p-\tau^p)^{\alpha-1}$9, defined through the cumulative Grünwald weights, together with a nilpotent lower-triangular matrix $u,v$0 whose powers reproduce Mittag–Leffler growth (Kumar et al., 2018).

A third representative formulation is given for convolution quadrature generated by $u,v$1. With

$u,v$2

and inverse kernel

$u,v$3

one has the exact inverse-kernel identity

$u,v$4

If

$u,v$5

then, under a time-step restriction,

$u,v$6

This formulation is tailored to convolution quadrature generated by generalized Newton–Gregory formulas (Yang et al., 2019).

3. Nonuniform meshes, complementary kernels, and asymptotic compatibility

The nonuniform-grid theory is built around discrete Caputo operators written in convolution form,

$u,v$7

with weights satisfying structural assumptions. A widely used framework imposes:

$u,v$8

a lower bound by local averages of the Caputo kernel,

$u,v$9

and a bounded local step ratio

$g$0

The central device is the complementary kernel $g$1, defined recursively so that

$g$2

If a nonnegative sequence satisfies

$g$3

then

$g$4

This formulation was designed to cover nonuniform L1, fast L1, and nonuniform Alikhanov/L2–$g$5 schemes (Liao et al., 2018).

The same complementary-kernel mechanism underlies analyses on irregular meshes for higher-order spatial discretizations. For the nonuniform L1 approximation

$g$6

with kernels satisfying positivity and monotonicity, complementary kernels $g$7 are constructed so that

$g$8

The associated discrete fractional Gronwall inequality then bounds sequences $g$9 satisfying

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$0

by a Mittag–Leffler factor involving the forcing terms and the complementary kernels. This framework was used on irregular meshes for a fourth-order compact solver for a fractional-in-time fourth-order diffusion equation (Zhong et al., 2019).

A more recent refinement is the asymptotically compatible formulation based on discrete convolution complementary kernels $$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$1. These are defined recursively by

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$2

and compared with the continuous weights

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$3

The key estimate is

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$4

which leads to the discrete inequality

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$5

As $$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$6,

$$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$7

so the inequality tends to the classical discrete Grönwall inequality (Yin et al., 2024).

A different nonuniform-grid line treats integral and differential discretizations through complete positivity and resolvents. For a kernel $$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$8 with inverse $$u(t) \le v(t) + g(t)\,p^{1-\alpha}\int_a^t \tau^{p-1}(t^p-\tau^p)^{\alpha-1}u(\tau)\,d\tau,$$9 satisfying

$v$0

one defines discrete resolvents $v$1 by

$v$2

This yields comparison principles and several Grönwall inequalities on arbitrary nonuniform meshes. For example, if

$v$3

then under a lower comparison with the continuous Caputo kernel one obtains

$v$4

while for

$v$5

one gets a Mittag–Leffler growth estimate under a local smallness condition on $v$6. The paper emphasizes that these results do not have any restrictions on the step size ratio (Feng et al., 2024).

4. Alternative discrete fractional calculi

The notion of discrete fractional Gronwall inequality is not confined to time discretizations of PDEs. In discrete Riemann–Liouville fractional calculus, the fractional sum

$v$7

and the fractional difference

$v$8

lead to solution kernels expressed through a discrete Mittag–Leffler-type function $v$9. A generalized Bernoulli inequality,

$$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$0

is then used as a building block for discrete fractional comparison and Gronwall-type bounds (Ferreira, 2017).

On time scales, the Caputo fractional delta operator produces a Gronwall theorem that simultaneously covers continuous and discrete cases. If

$$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$1

with $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$2 nonnegative, nondecreasing, and bounded, then

$$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$3

Specializing the time scale to $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$4 or $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$5 converts the delta integral into sums and yields a discrete fractional Gronwall inequality with kernels $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$6 (Pachpatte, 2019).

A further generalization appears in $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$7-fractional calculus on the geometric time scale

$$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$8

For the Caputo $$u(t) \le v(t)\,E_\alpha\big(g(t)\Gamma(\alpha)(t^p-a^p)^\alpha\big).$$9-fractional derivative, if

$V$0

then

$V$1

In special cases this series is written in terms of $V$2-Mittag–Leffler functions, and for $V$3 it reduces to a $V$4-exponential bound (Abdeljawad et al., 2013).

These frameworks show that the term “discrete fractional Gronwall inequality” covers both numerical-analysis estimates for approximate fractional derivatives and intrinsic comparison principles in discrete fractional calculus itself.

5. Numerical-analysis role: stability, convergence, and pointwise control

In numerical analysis, discrete fractional Gronwall inequalities are used in exactly the place occupied by classical discrete Gronwall for parabolic problems: they close energy estimates and convert nonlocal recursions into global-in-time stability and error bounds.

For nonlinear parabolic equations discretized by L1-Galerkin finite element methods, the L1 discrete fractional Gronwall inequality is applied to

$V$5

after proving

$V$6

The resulting estimate is

$V$7

and analogous higher-order linearized schemes satisfy

$V$8

The same paper states that the new inequality removes earlier small-time restrictions such as

$V$9

from nonlinear analyses (Li et al., 2016).

For the fractional Crank–Nicolson–Galerkin method based on backward Euler convolution quadrature, the discrete fractional Gronwall inequality is applied twice: first to prove a uniform $u\le v+Vu$0 bound

$u\le v+Vu$1

and then to the projected error $u\le v+Vu$2, leading to

$u\le v+Vu$3

In that setting the discrete inequality is driven by the Riemann–Liouville convolution quadrature operator and a discrete energy inequality

$u\le v+Vu$4

(Kumar et al., 2018).

For subdiffusion on nonuniform grids, the general complementary-kernel framework yields stability estimates of the form

$u\le v+Vu$5

and corresponding error bounds for fully discrete reaction–subdiffusion problems (Liao et al., 2018).

For a fourth-order compact solver for the fractional-in-time fourth-order diffusion equation, the discrete fractional Gronwall inequality on irregular meshes is combined with an error convolution structure. This yields

$u\le v+Vu$6

in $u\le v+Vu$7, and

$u\le v+Vu$8

in $u\le v+Vu$9 on graded meshes satisfying the stated mesh assumptions (Zhong et al., 2019).

The completely positive, resolvent-based theory has been applied to subdiffusion and time-fractional Allen–Cahn equations. For the subdiffusion problem, one obtains a uniform-in-time error bound

$q$00

and decay to the discrete steady state through Mittag–Leffler factors. For the time-fractional Allen–Cahn equation, one gets an $q$01-decay estimate of the form

$q$02

under the stated assumptions (Feng et al., 2024).

The asymptotically compatible DCC theory sharpens pointwise-in-time control on graded and quasi-graded meshes. If the error $q$03 satisfies

$q$04

then, on meshes $q$05, the bound

$q$06

is obtained (Yin et al., 2024).

6. Structural themes, limitations, and recurrent issues

Several structural themes recur throughout the literature. Positivity and monotonicity of the discrete memory weights are central. In the L1 theory,

$q$07

and the auxiliary kernels $q$08 satisfy $q$09 (Li et al., 2016). In the nonuniform Caputo framework, the assumptions

$q$10

and bounded local step ratio $q$11 are explicit hypotheses (Liao et al., 2018). In the nonuniform L1 theory for compact schemes, analogous assumptions are written as

$q$12

(Zhong et al., 2019).

A second theme is the appearance of inverse or complementary kernels. The L1 paper uses $q$13, the convolution-quadrature paper uses $q$14 and $q$15, the nonuniform-grid framework uses $q$16, and the asymptotically compatible theory uses DCC kernels $q$17 (Li et al., 2016, Yang et al., 2019, Liao et al., 2018, Yin et al., 2024). These kernels play the discrete role of the fractional integral and are what turn the nonlocal derivative inequality into an explicit sequence bound.

A third theme is that step-size conditions are common but not universal. The L1 inequality requires $q$18, the backward-Euler convolution-quadrature inequality requires $q$19, the general nonuniform Caputo theory imposes

$q$20

and the resolvent-based growth estimate assumes a local condition on $q$21 (Li et al., 2016, Kumar et al., 2018, Liao et al., 2018, Feng et al., 2024). By contrast, the resolvent-based comparison theory explicitly states that its Grönwall inequalities on nonuniform meshes do not have any restrictions on the step size ratio (Feng et al., 2024).

The recent literature also clarifies limitations of the prevailing energy framework. One paper proves that monotonicity of the discrete convolution kernels is not only sufficient but necessary for the quadratic inequality

$q$22

and notes that this excludes certain higher-order schemes, explicitly mentioning L1$q$23 and Caputo BDF$q$24, from direct treatment by the same approach (Yin et al., 2024). Relatedly, for Caputo BDF2-like schemes on nonuniform meshes, extending the complementary-kernel Grönwall theory remains an open problem because positivity and monotonicity of recombined kernels are not yet established on general meshes (Liao et al., 2018).

The literature therefore presents discrete fractional Gronwall inequality not as a single theorem, but as a scheme-dependent analytic framework. Its invariant content is the replacement of exponential comparison by Mittag–Leffler comparison for discrete operators with memory, together with the use of auxiliary inverse kernels, resolvents, or discrete fractional sums to control nonlocal recursions.