Tikhonov Guidance for Ill-posed Inverse Problems

• Tikhonov guidance is a framework of parameter selection, stopping rules, and iterative methods that stabilize Tikhonov regularization for solving noisy, ill-posed inverse problems.
• It employs range-relaxed rules to ensure the next residual remains within a feasible interval, balancing convergence speed with robustness against noise.
• The approach integrates nonstationary iterated schemes, discrepancy stopping, and Krylov-based implementations to optimize computational efficiency while maintaining accuracy.

Tikhonov guidance denotes the body of parameter-choice rules, stopping criteria, and iterative constructions that make Tikhonov regularization operational for ill-posed inverse problems with noisy data. In the linear Hilbert-space setting, it concerns equations $Ax=y$ with only $y^\delta$ available, and it includes single-shot Tikhonov, iterated and nonstationary iterated schemes, discrepancy-based rules, heuristic rules for unknown noise, and Krylov-projected implementations. A particularly explicit recent formulation is the range-relaxed nonstationary iterated Tikhonov method, which chooses iteration-dependent Lagrange multipliers so that the next residual lies in a feasible interval rather than at one prescribed value (Boiger et al., 2020).

1. Problem class and basic regularization model

The standard setting uses separable Hilbert spaces $X$ and $Y$, a bounded linear operator $A:X\to Y$, and noisy data $y^\delta\in Y$ satisfying

$\|y-y^\delta\|_Y \le \delta,$

with known noise level $\delta>0$. Exact solvability is assumed in the sense that there exists $x^\star\in X$ with $Ax^\star=y$, while ill-posedness means that $y^\delta$0 is either not defined or discontinuous. The distinguished exact solution is the minimal-norm solution $y^\delta$1, and an optional source condition of the form $y^\delta$2, $y^\delta$3, is used to discuss rates (Boiger et al., 2020).

The one-shot Tikhonov problem minimizes

$y^\delta$4

with normal equation

$y^\delta$5

This formulation stabilizes the inversion by balancing fidelity and norm control. In the same setting, residual level sets

$y^\delta$6

are closed and convex, satisfy $y^\delta$7 for $y^\delta$8, contain $y^\delta$9 whenever $X$0, and have nonempty interior if $X$1 (Boiger et al., 2020).

This geometric viewpoint is central. It turns parameter choice into the problem of determining how aggressively one should move an iterate toward a residual level set that is compatible with the noise floor. That interpretation is what underlies range-relaxed Tikhonov guidance.

2. Nonstationary iterated Tikhonov and the range-relaxed rule

Iterated Tikhonov replaces the one-shot penalty $X$2 by a proximal penalty around the current iterate: $X$3 Its optimality condition is

$X$4

If $X$5, one obtains stationary iterated Tikhonov; if $X$6 varies, one obtains nonstationary iterated Tikhonov. Writing $X$7, the update becomes

$X$8

so $X$9 acts as a Lagrange multiplier in a constrained projection step (Boiger et al., 2020).

A common a priori choice is geometric growth, $Y$0, equivalently $Y$1. The cited analysis identifies the usual trade-off: this can be fast, but it may be unstable if $Y$2 is too large or too slow if $Y$3 is too small. Exact a posteriori target-residual prescriptions avoid that rigidity, but they require solving a nonlinear scalar equation at every iteration (Boiger et al., 2020).

The range-relaxed rule addresses both issues by requiring only that the next residual remain in an admissible interval: $Y$4 Thus the new iterate is not forced to hit one exact residual target. Instead, it is accepted whenever its residual lies between the noise floor and a relaxed contraction of the current residual. Geometrically, $Y$5 is an orthogonal projection of $Y$6 onto some $Y$7 with

$Y$8

The corresponding constrained problem is

$Y$9

and its Lagrangian, for fixed $A:X\to Y$0, is

$A:X\to Y$1

For $A:X\to Y$2, the minimizer is

$A:X\to Y$3

with $A:X\to Y$4 (Boiger et al., 2020).

Defining

$A:X\to Y$5

the practical rule becomes

$A:X\to Y$6

A decisive structural fact is that $A:X\to Y$7 is strictly decreasing on $A:X\to Y$8, because

$A:X\to Y$9

Hence the admissible $y^\delta\in Y$0 form an interval $y^\delta\in Y$1 with $y^\delta\in Y$2. This interval-valued admissibility is the defining feature of range-relaxed guidance (Boiger et al., 2020).

3. Convergence, stability, and discrepancy stopping

The range-relaxed criterion implies a direct residual contraction law: $y^\delta\in Y$3 and therefore

$y^\delta\in Y$4

This gives exponential decay of the “distance to the noise floor” and immediately implies that Morozov discrepancy stopping,

$y^\delta\in Y$5

is finite. The stopping index obeys the explicit bound

$y^\delta\in Y$6

(Boiger et al., 2020).

The same analysis yields a monotonic error identity. Writing $y^\delta\in Y$7, one has

$y^\delta\in Y$8

Thus the squared error decreases strictly. Before discrepancy is met, the gain is bounded below by

$y^\delta\in Y$9

which identifies both primal movement and residual curvature as sources of descent (Boiger et al., 2020).

In the exact-data case $\|y-y^\delta\|_Y \le \delta,$0, the lower bound

$\|y-y^\delta\|_Y \le \delta,$1

implies $\|y-y^\delta\|_Y \le \delta,$2. By the Brill–Schock criterion for nonstationary iterated Tikhonov, this guarantees strong convergence of the iterates to a solution of $\|y-y^\delta\|_Y \le \delta,$3 (Boiger et al., 2020).

Under the source condition $\|y-y^\delta\|_Y \le \delta,$4, the cited work aligns the range-relaxed analysis with classical Tikhonov balance arguments. Single-shot Tikhonov obeys

$\|y-y^\delta\|_Y \le \delta,$5

and the range-relaxed nonstationary method is compatible with convergence $\|y-y^\delta\|_Y \le \delta,$6 as $\|y-y^\delta\|_Y \le \delta,$7, with order-optimal scaling recovered when the effective parameter at stopping satisfies the usual balance $\|y-y^\delta\|_Y \le \delta,$8 (Boiger et al., 2020).

4. Algorithmic realization and parameter heuristics

The implementable form of range-relaxed nonstationary iterated Tikhonov takes as input $\|y-y^\delta\|_Y \le \delta,$9, $\delta>0$0, $\delta>0$1, the noise level $\delta>0$2, an initial guess $\delta>0$3, a relaxation parameter $\delta>0$4, and a discrepancy factor $\delta>0$5. Typical values reported for the relaxation are $\delta>0$6, with $\delta>0$7 described as a good trade-off between fast residual decrease and robust feasibility; typical discrepancy factors are $\delta>0$8 in practical guidance, with $\delta>0$9 also discussed in relation to computational cost and accuracy (Boiger et al., 2020).

At iteration $x^\star\in X$0, one first checks discrepancy. If $x^\star\in X$1, the process stops. Otherwise, one defines the upper residual target

$x^\star\in X$2

and searches for $x^\star\in X$3 satisfying

$x^\star\in X$4

A practical initialization is the lower bound

$x^\star\in X$5

For $x^\star\in X$6, two heuristics are explicitly suggested: reusing $x^\star\in X$7, or extrapolating on the logarithmic scale,

$x^\star\in X$8

The inner solve is a monotone Newton-like iteration on $x^\star\in X$9: $Ax^\star=y$0 with $Ax^\star=y$1, doubling $Ax^\star=y$2 when $Ax^\star=y$3, and otherwise setting $Ax^\star=y$4. The inner loop stops as soon as

$Ax^\star=y$5

The reported behavior is that the inner Newton loop typically needs only a few steps, often $Ax^\star=y$6–$Ax^\star=y$7, because the residual map is monotone and exact target matching is unnecessary (Boiger et al., 2020).

Each evaluation of $Ax^\star=y$8 and $Ax^\star=y$9 requires solving

$y^\delta$00

equivalently $y^\delta$01 with $y^\delta$02. The cited implementation guidance recommends preconditioned CG, multigrid, FFT-based solvers for convolution operators, or direct factorizations, depending on the application class (Boiger et al., 2020).

The method’s practical advantages and limitations are sharply delimited. The feasible interval $y^\delta$03 protects against overshooting the residual target, unlike pure geometric growth $y^\delta$04, which may become unstable for large $y^\delta$05. At the same time, the method does not remove the cost of solving $y^\delta$06 systems, and if those systems dominate runtime or are poorly conditioned, preconditioning remains essential. Very large noise or very small $y^\delta$07 can narrow the feasible interval and increase inner iterations, while poor initial guesses can increase $y^\delta$08 (Boiger et al., 2020).

5. Empirical behavior in deblurring and inverse potential identification

The reported numerical study evaluates range-relaxed nonstationary iterated Tikhonov against geometric nonstationary iterated Tikhonov and an a posteriori Donatelli–Hanke variant in two distinct settings: image deblurring and a two-dimensional inverse potential problem. The metrics are relative reconstruction error $y^\delta$09, residual $y^\delta$10, and the number of linear systems solved as a proxy for cost. The central empirical findings are exponential residual decay consistent with theory, fewer total linear solves than the a posteriori Donatelli–Hanke method, and robustness in low-noise regimes (Boiger et al., 2020).

Application and noise	Setup	Total linear systems
Deblurring, $y^\delta$11	$y^\delta$12, $y^\delta$13, $y^\delta$14	gNIT 6; Donatelli–Hanke 15; rrNIT 7
Deblurring, $y^\delta$15	same	gNIT 17; Donatelli–Hanke 23; rrNIT 11
Deblurring, $y^\delta$16	same	gNIT 36; Donatelli–Hanke 43; rrNIT 16
Inverse potential, $y^\delta$17	$y^\delta$18, $y^\delta$19, $y^\delta$20	gNIT 6; Donatelli–Hanke 11; rrNIT 6
Inverse potential, $y^\delta$21	same	gNIT 10; Donatelli–Hanke 34; rrNIT 10
Inverse potential, $y^\delta$22	same	gNIT 13; Donatelli–Hanke 86; rrNIT 12

In the deblurring experiment, $y^\delta$23 is a space-invariant convolution with Gaussian point-spread function on a $y^\delta$24 image, implemented with FFT-based $y^\delta$25 and $y^\delta$26; the Gaussian point-spread function uses size $y^\delta$27 and $y^\delta$28. In the inverse potential problem, the forward PDE is

$y^\delta$29

with

$y^\delta$30

The unknown source has sharp gradients, so the problem probes whether the method remains effective beyond smooth-image benchmarks (Boiger et al., 2020).

Two empirical observations are especially relevant for parameter guidance. First, $y^\delta$31 tends to grow exponentially across iterations, so log-extrapolation for $y^\delta$32 substantially reduces inner Newton effort. Second, the gain over exact target-residual a posteriori rules becomes more pronounced at very low noise, because accepting any $y^\delta$33 inside a feasible interval is cheaper than solving a scalar nonlinear equation to high precision at every step (Boiger et al., 2020).

6. Position within the broader Tikhonov parameter-choice literature

Range-relaxed nonstationary iterated Tikhonov is one member of a larger family of Tikhonov guidance strategies. In Krylov-subspace settings, one influential approach embeds the discrepancy principle inside Arnoldi–Tikhonov by estimating the unknown noise norm through the GMRES residual plateau,

$y^\delta$34

and choosing $y^\delta$35 in reduced space so that

$y^\delta$36

thereby turning parameter selection into a low-dimensional update that reuses Arnoldi quantities (Gazzola et al., 2013). A complementary Morozov-based line of work formulates the Tikhonov solution and the regularization parameter as a single nonlinear system and solves it with a projected Newton direction computed in a Krylov subspace, coupled with an Armijo line search; this yields a globally convergent noise-constrained algorithm for large-scale problems (Cornelis et al., 2019).

When the data noise level is unknown but an operator perturbation bound $y^\delta$37 is available, semi-heuristic rules modify standard heuristic functionals such as heuristic discrepancy, Hanke–Raus, or quasi-optimality by subtracting compensators like $y^\delta$38 or $y^\delta$39, and by enforcing a lower search bound $y^\delta$40; these rules are designed precisely for the case in which operator error does not satisfy the same irregularity assumptions as data noise (Hämarik et al., 2018). Under white Gaussian noise, a different route is to minimize a lower bound of the predictive risk,

$y^\delta$41

whose minimizer depends only on the signal-to-noise ratio; when $y^\delta$42 and $y^\delta$43 are unknown, the iterative I-PRO scheme alternates between regularization-parameter minimization and signal-to-noise-ratio estimation (Benvenuto et al., 2019).

Heuristic rules remain important when no reliable noise bound is available, but they are not interchangeable. For quasi-optimality, one detailed analysis shows that the global minimizer can fail badly, whereas one proper local minimizer of the quasi-optimality function is always a good regularization parameter in the finite-dimensional standard Tikhonov setting (Raus et al., 2017). In models with several competing structural priors, Tikhonov guidance becomes vector-valued: multi-parameter Tikhonov regularization uses several penalties simultaneously and selects the parameter vector by discrepancy or balancing principles, with the balancing system

$y^\delta$44

as the central equilibrium condition (Ito et al., 2011).

The notion also extends beyond linear Hilbert-space inverse problems. For nonlinear problems in Banach spaces, Bregman-distance penalization yields stability, convergence, and rates in Bregman distances, and supports an iterated Tikhonov method with discrepancy stopping (Bleyer et al., 2020). In Banach-space oversmoothing, where the exact solution lies outside the penalty space, convergence can still be proved in weaker norms, with rates of the form

$y^\delta$45

under sectorial-operator and interpolation-scale assumptions (Chen et al., 2020). In Krylov hybridization, CG–Tikhonov can be interpreted as a filtration of the CG Lanczos vectors, with Lanczos filters $y^\delta$46 damping later, noise-amplifying components more strongly as $y^\delta$47 (Gerth et al., 29 May 2025), and the iterated Golub–Kahan–Tikhonov method achieves

$y^\delta$48

under the source conditions and balancing strategy stated in its analysis (Bianchi et al., 16 Jul 2025).

Taken together, these developments indicate that Tikhonov guidance is not a single rule but a structured design problem. The relevant choice depends on what is known about the noise, whether the forward operator is perturbed, whether the problem is reduced in a Krylov subspace, whether multiple penalties are used, and whether the underlying geometry is Hilbertian, Banach, or projection-based. Within that landscape, the range-relaxed criterion stands out for replacing exact target matching by a guaranteed feasible interval, thereby linking geometric projection, robust convergence, and low inner-solve complexity in a particularly direct way (Boiger et al., 2020).