Monotonicity Method in Inverse Problems

• Monotonicity method is a family of techniques using order relations to infer structure and guarantee uniqueness and convergence in various applications.
• It employs operator inequalities, coercive-plus-compact decompositions, and localized potentials to achieve robust reconstruction in inverse problems and scattering.
• The approach extends to PDE analysis, optimization, and decision theory, underpinning algorithmic stability and precise numerical methods.

“Monotonicity method” is not a single technique with one fixed definition; it is a family of methods that exploit an order relation to infer structure, prove uniqueness or convergence, or exclude counterintuitive behavior. In inverse problems, the order is typically between boundary or far-field operators and artificial test operators; in PDE analysis it is often a monotone or almost-monotone frequency functional; in optimization it appears as enforced decrease of objective values or merit functions; in numerical analysis it is tied to inverse positivity, discrete maximum principles, or absolute monotonicity; and in decision theory and statistics it can be formulated as an axiom or hypothesis on how outputs should respond to favorable perturbations (Garde et al., 2017, Daimon et al., 2019, Felli et al., 2011, Nishimura et al., 2022, Csató et al., 2019).

1. General architecture and order structures

A recurring pattern in the inverse-problem literature is to encode the unknown object into an operator and then compare that operator with a synthetic one attached to a trial set. The comparison is expressed either in the positive-semidefinite sense, as in the Loewner order $A \preceq B$, or in a weakened spectral order that allows finitely many exceptional directions. In scattering, the standard notation is

$A \le_{\mathrm{fin}} B,$

meaning that $B-A$ has only finitely many negative eigenvalues; equivalently, positivity holds on the orthogonal complement of a finite-dimensional subspace. This order is central in acoustic crack scattering, mixed obstacle scattering, rigid elastic obstacle scattering, and elastic scattering on unbounded domains (Daimon et al., 2019, Furuya, 2021, Bai et al., 5 Jun 2025, Harrach et al., 4 Feb 2026).

The abstract mechanism is usually a coercive-plus-compact decomposition. In the acoustic setting, a far-field operator is written in factored form such as $F=GTG^*$, and one shows that $\Re T=C+K$ with $C$ positive coercive and $K$ compact. This produces one-sided operator inequalities for trial domains contained in the target. The converse direction is then obtained by range-separation arguments or localized potentials, which construct waves concentrated in a region where a false trial set protrudes outside the true object (Furuya, 2021). This suggests a broad methodological schema: monotonicity yields the “inside” implication, while localization or disjoint-range arguments yield sharpness.

Outside inverse problems, the ordered quantity changes but the logic remains comparable. Almgren-type methods study the near-monotonicity of a frequency function $\mathcal N(r)=D(r)/H(r)$ as $r\to 0^+$ to classify blow-up behavior at singular boundary points (Felli et al., 2011). Multiobjective optimization enforces weak monotonicity of objective values by accepting an accelerated proximal point only when at least one objective does not increase (Nishimura et al., 2022). Time integrators quantify monotonicity by the radius of absolute monotonicity, which is also the SSP coefficient for irreducible Runge–Kutta methods (Bonaventura et al., 2015). In pairwise-comparison ranking, monotonicity becomes an axiom requiring that strengthening a preference for alternative $i$ over $A \le_{\mathrm{fin}} B,$0 must not lower $A \le_{\mathrm{fin}} B,$1’s rank or normalized weight (Csató et al., 2019).

2. Bounded-domain inverse problems

In electrical impedance tomography, the monotonicity method is built on the Neumann-to-Dirichlet map $A \le_{\mathrm{fin}} B,$2 for the conductivity equation and on the basic order relation

$A \le_{\mathrm{fin}} B,$3

For irregular indefinite inclusions, the conductivity is written as $A \le_{\mathrm{fin}} B,$4, with open disjoint positive and negative parts. The reconstruction target is not necessarily the exact inclusion but its outer support $A \le_{\mathrm{fin}} B,$5, the smallest closed superset with connected complement. The method tests admissible closed sets $A \le_{\mathrm{fin}} B,$6 by nonlinear and linearized operator inequalities and proves that, under a weak unique continuation property, the intersection of all accepted $A \le_{\mathrm{fin}} B,$7 equals $A \le_{\mathrm{fin}} B,$8. With the additional separation condition $A \le_{\mathrm{fin}} B,$9, the outer supports of the positive and negative parts can be reconstructed independently. The same paper develops regularization for approximate models, including the Complete Electrode Model, proves asymptotic exactness as modeling error and noise vanish, and introduces a peeling-type algorithm that shrinks a large initial candidate by boundary-pixel removal (Garde et al., 2017).

A closely related bounded-domain variant combines monotonicity with shape optimization. In an EIT inclusion problem with

$B-A$0

the linearized monotonicity test

$B-A$1

is used to generate a coarse inclusion indicator on a pixel partition. A regularized cellwise reconstruction is then simplified geometrically and used as the initial level-set for a Kohn–Vogelius shape optimization. This hybridization is motivated by the observation that level-set methods are sensitive to initialization, whereas monotonicity tests are global but coarse. In the reported experiments, the monotonicity initializer improves reliability, with level-set convergence in about $B-A$2 iterations in free-noise cases and $B-A$3 or $B-A$4 iterations for the stated noisy cases (Harrach et al., 27 Jan 2025).

In linear isotropic elasticity, the analogue is the Neumann-to-Dirichlet operator $B-A$5 for the Lamé system. The key inequality states that increasing $B-A$6 and $B-A$7 decreases $B-A$8. This supports two reconstruction schemes. The standard monotonicity method compares measured data with trial NtD operators $B-A$9. The linearized method replaces these forward solves by the Fréchet derivative $F=GTG^*$0, leading to a much cheaper offline stage. Both methods reconstruct the outer support of the inclusion via operator inequalities justified by localized potentials. Numerically, the standard method yields slightly better reconstructions, while the linearized method is dramatically faster in preprocessing; for the reported $F=GTG^*$1-block test, the standard offline phase takes about $F=GTG^*$2 days $F=GTG^*$3 h $F=GTG^*$4 min, versus $F=GTG^*$5 h $F=GTG^*$6 min $F=GTG^*$7 s for the linearized method (Eberle et al., 2020).

Time-harmonic elasticity on bounded domains extends the same logic to the Navier equation with density. For coefficient triples $F=GTG^*$8, the monotonicity inequality holds only modulo a finite-dimensional exceptional space $F=GTG^*$9, reflecting the frequency-dependent spectral defect. Under the sign pattern $\Re T=C+K$0 increased and $\Re T=C+K$1 decreased relative to the background, inclusion tests are obtained by comparing the measured NtD map with trial maps and counting negative eigenvalues. A later linearized development around a homogeneous background proves Fréchet differentiability of the NtD map and shows that the derivative-based test can do more than recover the combined support: when $\Re T=C+K$2, it can partially separate the support of $\Re T=C+K$3 from the rest of the inhomogeneity, using zero-divergence localized solutions in the constant-coefficient background (Eberle-Blick et al., 2023, Eberle-Blick et al., 2024).

3. Far-field monotonicity in scattering theory

For inverse acoustic crack scattering, the unknown is a smooth open arc $\Re T=C+K$4, and the data are encoded by the far-field operator $\Re T=C+K$5. The monotonicity method uses the self-adjoint operator $\Re T=C+K$6 and two complementary test operators. For a smooth open test arc $\Re T=C+K$7, Theorem 1.1 states

$\Re T=C+K$8

For a bounded open set $\Re T=C+K$9, Theorem 1.2 states

$C$0

The proofs rely on the factorization $C$1, open-arc trace spaces, coercivity of the single-layer operator at imaginary wavenumber, and range-separation arguments based on Rellich’s lemma and unique continuation. The implemented indicator counts the negative eigenvalues of $C$2; the reported numerics reconstruct the crack geometry from short vertical or horizontal line segments with $C$3, $C$4, $C$5, and $C$6, and the reconstruction is stated to work independently of the direction of the test segment (Daimon et al., 2019).

A general abstract theorem for the monotonicity method in acoustic scattering places these arguments in a unified functional-analytic setting. Given factorized operators $C$7 and $C$8, coercivity of $C$9 up to compact perturbations yields $K$0 under a compact comparison map $K$1, while infinite-dimensional disjointness of appropriate ranges forces failure of the inequality in the converse direction. The mixed version treats block operators with one positive-coercive and one negative-coercive diagonal block, which is precisely the structure arising for mixed boundary conditions. Compared with the factorization method, the monotonicity theorem works under weaker a priori assumptions, uses only $K$2, avoids positivity assumptions on the imaginary part, and directly handles mixed obstacles and mixed cracks without masking domains or auxiliary closed curves (Furuya, 2021).

In rigid elastic obstacle scattering, the same strategy is transferred to the two-dimensional Navier equation. The far-field operator satisfies

$K$3

with $K$4 the elastic single-layer operator and $K$5 the data-to-pattern map. For a probe domain $K$6, the central comparison is made through $K$7. If $K$8, then $K$9; if $\mathcal N(r)=D(r)/H(r)$0, localized wave functions yield the strict failure of this ordering. The method is non-iterative and sampling-based: one tests many domains $\mathcal N(r)=D(r)/H(r)$1 and inspects the number of negative eigenvalues of $\mathcal N(r)=D(r)/H(r)$2. The paper emphasizes that reconstruction does not require an initial guess and is not driven by nonlinear parameter fitting (Bai et al., 5 Jun 2025).

A further extension addresses penetrable elastic inhomogeneities in an unbounded homogeneous background. Here the unknown support is

$\mathcal N(r)=D(r)/H(r)$3

and the measured data are the far-field operator $\mathcal N(r)=D(r)/H(r)$4 for elastic Herglotz incident waves. The central monotonicity identity is written for $\mathcal N(r)=D(r)/H(r)$5, where $\mathcal N(r)=D(r)/H(r)$6 is the unitary scattering operator. Under the sign pattern $\mathcal N(r)=D(r)/H(r)$7, $\mathcal N(r)=D(r)/H(r)$8, and $\mathcal N(r)=D(r)/H(r)$9, the operator is positive modulo finitely many negative eigenvalues. Combined with localized potentials for the Navier equation, this yields support characterization for compactly supported perturbations of the Lamé parameters and density in an exterior scattering setting (Harrach et al., 4 Feb 2026).

4. PDE-analytic monotonicity and weak-solution theory

In elliptic PDE analysis, the monotonicity method can mean an Almgren-type frequency argument. For semilinear elliptic equations near a conical boundary point, the domain is first straightened so that the boundary becomes a lower-order perturbation of a cone. One then defines a renormalized energy $r\to 0^+$0, a renormalized boundary mass $r\to 0^+$1, and the frequency

$r\to 0^+$2

Using differential identities, a Pohozaev-type formula, Hardy-type inequalities, and the fact that the boundary profile is sufficiently close to a cone, the derivative of $r\to 0^+$3 is shown to consist of nonnegative terms plus integrable errors. Hence $r\to 0^+$4 has a finite limit $r\to 0^+$5 as $r\to 0^+$6. Blow-up sequences converge to a homogeneous solution $r\to 0^+$7 on the limiting cone, where $r\to 0^+$8 is an eigenfunction of the angular operator $r\to 0^+$9. This yields a precise asymptotic classification at the corner and excludes logarithmic corrections in the leading term under the stated geometric assumptions (Felli et al., 2011).

In mean-field games, monotonicity is formulated directly at the operator level. For the stationary first-order system on $i$0,

$i$1

the operator is monotone in $i$2 when $i$3 is convex and $i$4 is increasing. Strict monotonicity yields uniqueness, while existence is obtained by monotone regularization and Minty’s method. The weak formulation is the variational inequality

$i$5

and the regularized operators add higher-order coercive terms without destroying monotonicity. The chapter develops this program for stationary and time-dependent MFGs, proving existence of weak solutions under the stated structural assumptions and emphasizing that monotonicity provides both a uniqueness principle and a compactness-compatible existence theory (Ferreira et al., 27 Feb 2025).

5. Algorithmic monotonicity in optimization and operator splitting

For multiobjective composite optimization,

$i$6

a monotone accelerated proximal gradient scheme generalizes Beck–Teboulle’s MFISTA. The nonmonotone multiobjective FISTA step computes $i$7; the monotone variant accepts $i$8 only if

$i$9

that is, if at least one objective does not increase. Otherwise it keeps $A \le_{\mathrm{fin}} B,$00. This “weak monotonicity rule” is less restrictive than requiring all objectives to decrease. The method preserves the accelerated rate

$A \le_{\mathrm{fin}} B,$01

keeps every objective bounded by its initial value, and converges globally to weak Pareto points under the stated assumptions. In the reported image-deblurring example, ordinary FISTA diverged considerably, Strong-MFISTA failed to converge, and Weak-MFISTA converged in few iterations (Nishimura et al., 2022).

A different extension appears in the monotonicity-of-pairs framework for non-monotone inclusions $A \le_{\mathrm{fin}} B,$02. Instead of requiring $A \le_{\mathrm{fin}} B,$03 itself to be monotone, one constructs an auxiliary map $A \le_{\mathrm{fin}} B,$04 such that the pair $A \le_{\mathrm{fin}} B,$05 is monotone or strongly monotone. The associated warped resolvent is

$A \le_{\mathrm{fin}} B,$06

and the proposed GIPPA iteration uses inertial extrapolation

$A \le_{\mathrm{fin}} B,$07

Under the stated assumptions on $A \le_{\mathrm{fin}} B,$08, $A \le_{\mathrm{fin}} B,$09, $A \le_{\mathrm{fin}} B,$10, and $A \le_{\mathrm{fin}} B,$11, the method yields residual convergence, weak convergence, and, under strong monotonicity, strong and linear convergence. The paper also gives constructive recipes for $A \le_{\mathrm{fin}} B,$12 in linear and locally nonlinear settings, including the choice $A \le_{\mathrm{fin}} B,$13 near a solution (Le et al., 19 Jan 2026).

Monotonicity can also be restored in a black-box sense. Given oracle access to a possibly non-monotone function $A \le_{\mathrm{fin}} B,$14, a meta-algorithm constructs a feasible monotone function $A \le_{\mathrm{fin}} B,$15 satisfying

$A \le_{\mathrm{fin}} B,$16

The reported query complexity is logarithmic in $A \le_{\mathrm{fin}} B,$17 and exponential in dimension $A \le_{\mathrm{fin}} B,$18, and a lower bound shows that the exponential dependence on $A \le_{\mathrm{fin}} B,$19 is necessary. For the weaker requirement of $A \le_{\mathrm{fin}} B,$20-marginal monotonicity, the dependence improves to one that is polynomial in $A \le_{\mathrm{fin}} B,$21 and exponential only in $A \le_{\mathrm{fin}} B,$22 (Gergatsouli et al., 2020).

6. Discrete monotonicity, inverse positivity, and time integration

For high-order discretizations of the Laplacian, monotonicity means inverse positivity: $A \le_{\mathrm{fin}} B,$23 entrywise. This implies a discrete maximum principle. In two dimensions, the $A \le_{\mathrm{fin}} B,$24 spectral element method on a uniform rectangular mesh is proved monotone by factoring the stiffness matrix into a product of four M-matrices. The argument uses Lorenz’s monotonicity criterion, auxiliary matrices $A \le_{\mathrm{fin}} B,$25, and iterative factorizations showing

$A \le_{\mathrm{fin}} B,$26

Because each factor is an M-matrix, $A \le_{\mathrm{fin}} B,$27, and the scheme inherits a discrete maximum principle. The paper also interprets the method as a fifth-order accurate finite difference scheme and states that, to the authors’ knowledge, it is the first monotone two-dimensional scheme of that accuracy beyond fourth order (Cross et al., 2020).

For the $A \le_{\mathrm{fin}} B,$28 spectral element method on quasi-uniform rectangular meshes, monotonicity is again proved through a matrix factorization argument, but the key technical advance is a relaxed Lorenz condition. Instead of using the original diagonal $A \le_{\mathrm{fin}} B,$29, the proof introduces a larger diagonal $A \le_{\mathrm{fin}} B,$30, which relaxes the mesh restrictions needed to dominate the positive off-diagonal coefficients. The final theorem gives explicit local mesh constraints and a global sufficient quasi-uniformity condition

$A \le_{\mathrm{fin}} B,$31

for any two mesh sizes (Cross et al., 2023).

In time integration, monotonicity is expressed through absolute monotonicity and SSP theory. For the two-stage DIRK family TR-BDF2, the absolute monotonicity radius is

$A \le_{\mathrm{fin}} B,$32

and this is maximized at

$A \le_{\mathrm{fin}} B,$33

which is also the L-stable parameter value. At that point $A \le_{\mathrm{fin}} B,$34. Because higher-order methods cannot be unconditionally monotone, the paper proposes two hybrid variants that revert to implicit Euler when a monotonicity sensor is triggered. The blended and partitioned strategies preserve L-stability while sacrificing order when necessary. In the reported advection and conservation-law tests, the hybrid variants remain TVD at high CFL numbers where standard conditionally monotone methods develop oscillations (Bonaventura et al., 2015).

A complementary line studies perturbed or downwind Runge–Kutta methods. For a given RK method $A \le_{\mathrm{fin}} B,$35, one introduces a second operator $A \le_{\mathrm{fin}} B,$36 and seeks a perturbation $A \le_{\mathrm{fin}} B,$37 that maximizes the perturbed radius of absolute monotonicity $A \le_{\mathrm{fin}} B,$38. A general theorem states that for any explicit, diagonally implicit, or fully implicit RK method, there exists a perturbation in the same structural class with positive radius even if the original method has $A \le_{\mathrm{fin}} B,$39. The paper derives upper bounds, gives LP-based and splitting algorithms for explicit methods, and reports optimal perturbed coefficients for many classical schemes, including positive-radius perturbations of classical RK4 (Higueras et al., 2015).

7. Axiomatic, statistical, and interpretive uses

In pairwise-comparison theory, monotonicity is an axiom about how priorities should react when a decision-maker strengthens one comparison entry. For a multiplicative reciprocal matrix $A \le_{\mathrm{fin}} B,$40, rank monotonicity requires that increasing $A \le_{\mathrm{fin}} B,$41 must not make alternative $A \le_{\mathrm{fin}} B,$42 fall below any alternative it previously weakly dominated, while weight monotonicity requires that the normalized weight of $A \le_{\mathrm{fin}} B,$43 must not decrease. The row geometric mean method and the column sum method satisfy both axioms, but the eigenvector method does not. A six-by-six counterexample shows rank reversal after increasing a favorable entry, and the same example violates weight monotonicity. Simulations against Saaty’s inconsistency ratio show that violations become more frequent as inconsistency increases, but no rank-monotonicity violations were found for nearly consistent matrices with $A \le_{\mathrm{fin}} B,$44, and even for heavily inconsistent matrices the violations remained relatively rare. The paper concludes that users of AHP should not assume that increasing a preference judgment always benefits the favored alternative under the eigenvector method (Csató et al., 2019).

In econometrics, monotonicity is a null hypothesis on a regression function rather than an operator comparison. For the model $A \le_{\mathrm{fin}} B,$45, the paper tests whether $A \le_{\mathrm{fin}} B,$46 is nondecreasing using a maximized studentized weighted pairwise U-statistic

$A \le_{\mathrm{fin}} B,$47

Bootstrap critical values are combined with one-step and step-down selection algorithms to reduce conservativeness, and the procedure is extended to settings with multiple covariates, partially linear structure, endogeneity, and sample selection. The tests are proved to have correct asymptotic size and to be asymptotically nonconservative; for regression functions with Lipschitz-continuous first derivatives, the framework attains the best attainable rate of uniform consistency while adapting to unknown smoothness (Chetverikov, 2012).

The literature therefore does not use a single universal definition of the monotonicity method. In some areas it is an inclusion principle derived from ordered operators; in others it is a monotone frequency formula, an inverse-positivity criterion, an acceptance rule for accelerated algorithms, an SSP radius, or a response axiom for rankings and statistical models. What these uses share is not one formula but one methodological stance: order information is elevated from a qualitative intuition to a rigorous analytical device, and the success of the method depends on proving that the chosen order is both physically or decision-theoretically meaningful and sharp enough to distinguish true structure from admissible but false alternatives.