Relaxed Compensator: Theory & Applications

• Relaxed Compensator is a versatile concept that modifies traditional compensators through relaxation techniques like averaging, smoothing, and partial resets.
• In control and imaging, its application – such as partial resets in reset control and total variation regularization in IMRT – leads to improved performance metrics and system robustness.
• Theoretical frameworks demonstrate that relaxed compensators enable predictable construction under weakened regularity assumptions, facilitating strong convergence and stability in complex models.

Searching arXiv for papers and exact topic usage to ground the article. arxiv_search query: "relaxed compensator" Across the cited literature, the expression relaxed compensator does not denote a single canonical object. It designates several distinct constructions in which a compensating process, device, or parameter is modified by relaxation, averaging, smoothing, or self-verification. In stochastic-process theory, it is the predictable compensator of a discrete-time Hawkes process constructed under minimal regularity; in G-stochastic control, it is the product-measure compensator associated with a relaxed Poisson measure; in reset control, it refers to partial reset through a reset matrix $A_\rho=\gamma$; in preclinical IMRT, it is a smoother 3D-printed thickness map obtained by total variation regularization; in optical metrology, it is an autocollimating null-corrector whose design reduces alignment ambiguity; in relaxation limits for traffic-flow PDEs, the term describes the joint action of relaxation and compensated compactness; and in signal detection under unknown background, it is a scalar parameter $\delta$ that corrects background misspecification along the signal direction (Sarma et al., 2024, Gherbal et al., 2021, Dastjerdi et al., 2020, Liu et al., 2021, Terebizh, 2014, Juajibioy et al., 2013, Banerjee et al., 19 May 2026).

1. Terminological scope

The literature uses the term in several technically unrelated ways. The common feature is not a shared formula, but a shared role: each construction compensates for a source of difficulty while weakening a stronger requirement such as exact model specification, strict controls, abrupt modulation, or fragile alignment.

Context	Compensator object	Relaxed aspect
Discrete-time Hawkes process	$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$	Constructed by Doob’s decomposition without requiring stationarity or an immigration–birth representation
G-stochastic control with jumps	$\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$	Strict controls are replaced by measure-valued relaxed controls
Reset control / CgLp	Reset matrix $A_\rho=\gamma$	Partial reset with $-1<\gamma\le 1$
Small-animal compensator IMRT	3D-printed thickness map $z$	Total variation regularization reduces abrupt spatial variation
Aspheric optical testing	Autocollimating null-corrector	Independent self-test and optical distance setting reduce practical alignment difficulty
Relaxation limit of Aw–Rascle	Relaxation plus compensated compactness	Hyperbolic oscillations are controlled by source damping and diffusion
Signal detection	Scalar $\delta$ in the $L^2(G)$ expansion of $f_b/g$	Background uncertainty is reduced to a single parameter

A recurrent misconception is to treat relaxed compensator as a standardized term across fields. The cited works do not support that reading. Another misconception is that “relaxed” necessarily means “looser tolerances.” In the autocollimating optical design, for example, the tolerances remain tight; what is relaxed is the practical difficulty of verification and positioning, not the numerical tolerance budget (Terebizh, 2014).

2. Predictable compensators under relaxed probabilistic assumptions

In the discrete-time Hawkes setting, the compensator arises from a binary self-exciting sequence $\delta$0 with conditional probabilities determined by a positive kernel $\delta$1. Defining the counting process $\delta$2 and filtration $\delta$3, the discrete-time intensity is

$\delta$4

The process $\delta$5 is a submartingale, and Doob’s decomposition yields

$\delta$6

with

$\delta$7

Thus $\delta$8 is the unique increasing predictable process with $\delta$9 a martingale. The “relaxed” aspect is that this construction is obtained purely from submartingale structure and predictability/integrability, without requiring stationarity, mixing, or an immigration–birth representation (Sarma et al., 2024).

Under the kernel assumptions

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$0

the long-run mean is

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$1

The strong law gives $$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$2 almost surely, and the compensator inherits the same limit: $$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$3 almost surely and in probability. The central limit theorem for the compensator is

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$4

where

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$5

The same paper also studies the scaled log-MGF

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$6

establishing, for $$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$7,

$$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$8

and for $$\Lambda_n=\sum_{i=1}^n E(\xi_i\mid\mathcal F_{i-1})$$9,

$\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$0

with $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$1 strictly decreasing in $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$2 for $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$3 and converging pointwise to a limit $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$4 (Sarma et al., 2024).

A different probabilistic use appears in G-stochastic control with controlled jumps. There, strict controls $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$5 are replaced by relaxed controls

$\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$6

where $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$7 takes values in $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$8. The jump mechanism is lifted to a relaxed counting measure on $\nu(dt,d\theta,da)=\rho_t(da)\,v(d\theta)\,dt$9, and the compensator is the product measure

$A_\rho=\gamma$0

The compensated relaxed measure is

$A_\rho=\gamma$1

The relaxed state equation averages the coefficients with respect to $A_\rho=\gamma$2, while the G-chattering lemma gives quasi-sure approximation of relaxed controls by strict controls. Under assumptions (A1)–(A4), the strict and relaxed problems have the same value function, and the relaxed maximum principle is formulated through the Hamiltonian and an adjoint G-BSDE with jumps (Gherbal et al., 2021).

Taken together, these two lines of work preserve the classical compensator idea as the predictable finite-variation part in a martingale structure, while relaxing either the regularity assumptions on the point process or the admissible-control class.

3. Partial-reset compensators in nonlinear loop shaping

In reset control, a relaxed compensator is a partial reset element. The reset law is

$A_\rho=\gamma$3

with $A_\rho=\gamma$4. The case $A_\rho=\gamma$5 is full reset; nonzero $A_\rho=\gamma$6 gives partial reset. The paper on CgLp compensators uses this notion in a frequency-domain tuning framework that explicitly accounts for high-order harmonics through pseudo-sensitivities rather than relying only on the classical describing function (Dastjerdi et al., 2020).

The first-order CgLp element is constructed as a reset FORE cascaded with a linear lead filter,

$A_\rho=\gamma$7

with $A_\rho=\gamma$8 and $A_\rho=\gamma$9 so that the magnitude remains approximately constant while the phase exhibits lead. The tuning method combines loop-shaping constraints at the target crossover $-1<\gamma\le 1$0 with a modulus-margin constraint based on pseudo-sensitivity $-1<\gamma\le 1$1, an iso-damping phase-slope constraint, and the $-1<\gamma\le 1$2 stability condition. The tuning objective is

$-1<\gamma\le 1$3

in dB.

The reported experiment uses a precision positioning stage (“Spider”) with plant model

$-1<\gamma\le 1$4

and specifications $-1<\gamma\le 1$5 Hz, phase margin $-1<\gamma\le 1$6, modulus margin $-1<\gamma\le 1$7 dB, $-1<\gamma\le 1$8, and $-1<\gamma\le 1$9. The tuned CgLp-based controller outperformed the linear PID comparator: step responses had the same rise time but less overshoot and shorter settling time; triangular-reference tracking at amplitude $z$0m improved by approximately $z$1; white-noise rejection improved by approximately $z$2; and disturbance rejection for $z$3A improved by approximately $z$4 (Dastjerdi et al., 2020).

Here, “relaxed” does not mean a smoother physical surface or a stochastic compensator. It means that the internal reset map is relaxed from full reset to fractional reset. The resulting nonlinearity is then managed through pseudo-sensitivities that account for high-order harmonics.

4. Relaxed physical compensators in radiation therapy and optical metrology

In small-animal 3D-printed compensator IMRT, a relaxed compensator is a thickness pattern with reduced spatial variation. The treatment fluence is represented by beamlet intensities $z$5, mapped to dose by a dose-influence matrix $z$6 and regularized by the anisotropic total variation penalty

$z$7

The optimization problem is

$z$8

A smooth surrogate for the $z$9 term is used for LBFGS optimization. Beam transmission through copper-doped PLA follows Beer–Lambert attenuation,

$\delta$0

and the optimized fluence is converted to thickness using the measured transmission function $\delta$1. The study used an Xstrahl SARRP with five beams at gantry angles $\delta$2, $\delta$3, $\delta$4, $\delta$5, and $\delta$6, a beamlet resolution of $\delta$7 mm at SAD, a $\delta$8 mm copper-PLA base layer, and a $\delta$9 mm maximum thickness in the modulated region (Liu et al., 2021).

Total variation regularization reduced intensity TV from $L^2(G)$0 to $L^2(G)$1 and thickness TV from $L^2(G)$2 mm to $L^2(G)$3 mm. The sum of compensator thicknesses decreased from $L^2(G)$4 mm to $L^2(G)$5 mm, and exposure time from $L^2(G)$6 s to $L^2(G)$7 s. Dosimetric quality remained comparable: PTV $L^2(G)$8 was $L^2(G)$9 versus $f_b/g$0, PTV $f_b/g$1 was $f_b/g$2 versus $f_b/g$3, and the hotspot metric $f_b/g$4 changed from $f_b/g$5 to $f_b/g$6. Composite gamma passing rate at $f_b/g$7 mm improved from $f_b/g$8 to $f_b/g$9 (Liu et al., 2021).

In aspheric optical testing, the compensator is a null-corrector. The autocollimating design consists of three spherical lenses made of the same glass, with the rear surface of the third lens flat. In autocollimation mode A, a point source at $\delta$00 mm sends light through the compensator, reflects from a flat, and returns to form a diffraction-limited image at the source. In control mode C, the same compensator is used to null-test a fast aspheric surface, exemplified by the nominal Hubble Space Telescope primary with $\delta$01 mm, $\delta$02 mm, $\delta$03, and surface speed approximately $\delta$04 (Terebizh, 2014).

The key geometric relation for the asphere is the aberration of normals,

$\delta$05

and the paper gives the axial setting accuracy estimate

$\delta$06

with $\delta$07 in radians and $\delta$08 in microns. In mode A, RMS wavefront error is $\delta$09 at $\delta$10m, the Airy diameter is approximately $\delta$11m, and $\delta$12 rad gives $\delta$13m. In mode C, RMS wavefront error is $\delta$14, the Airy diameter is approximately $\delta$15m, and $\delta$16 rad gives $\delta$17m. The tolerances required for $\delta$18 RMS remain tight: radii $\delta$19–$\delta$20 mm, thicknesses and surface decenters $\delta$21–$\delta$22 mm, element decenters of several microns, transverse displacement $\delta$23m, $\delta$24, and Abbe-number tolerance $\delta$25 (Terebizh, 2014).

This optical usage makes clear that “relaxed” refers to procedural robustness. The autocollimating design embeds independent verification of the assembled compensator and optical setting of key distances, thereby reducing the risk of wrong-null operation without loosening the underlying optical tolerances.

5. Relaxation and compensated compactness in continuum models

For the Aw–Rascle system with stiff relaxation and domain diffusion, the phrase is used in a different, methodological sense. The model is

$\delta$26

with parabolic approximation

$\delta$27

The analysis uses invariant regions to obtain uniform bounds and positivity, and compensated compactness to pass to the limit in nonlinear fluxes. The main convergence result states that, as $\delta$28 with $\delta$29, a subsequence converges almost everywhere to an equilibrium state satisfying

$\delta$30

and

$\delta$31

Thus the second-order system collapses to a first-order LWR-type law (Juajibioy et al., 2013).

The invariant region is

$\delta$32

with Riemann invariants

$\delta$33

Uniform estimates are derived from the functional

$\delta$34

leading to the key inequality

$\delta$35

Under $\delta$36, this yields local $\delta$37 bounds on $\delta$38, $\delta$39, and $\delta$40, which feed into entropy compactness and Murat’s lemma (Juajibioy et al., 2013).

In this setting, “relaxed compensator” is not a compensator in the martingale sense. It names the way the relaxation source term, diffusion, and compensated compactness jointly compensate for the lack of strong compactness in the hyperbolic system. The expression is therefore methodological rather than object-level.

6. Scalar compensators in inference under unknown background

A further use appears in statistical signal detection. The observed density is modeled as

$\delta$41

with test $\delta$42 versus $\delta$43. Choosing a proposal background density $\delta$44, one defines

$\delta$45

Expanding the unknown background ratio in an orthonormal basis $\delta$46 of $\delta$47 gives

$\delta$48

The scalar coefficient

$\delta$49

is the compensator. It is the projection of $\delta$50 onto the signal direction and is sufficient, together with

$\delta$51

for inference on $\delta$52 through

$\delta$53

The paper emphasizes that estimating the full background distribution is unnecessary for inference on signal intensity; it suffices to estimate $\delta$54 (Banerjee et al., 19 May 2026).

With a background-only sample $\delta$55, the estimators are

$\delta$56

and

$\delta$57

This supports Wald tests and confidence intervals with explicit plug-in variance formulas. Without a background-only sample, $\delta$58 is not identifiable, so the paper targets the conservative quantity

$\delta$59

under the sufficient condition $\delta$60. To enforce that condition, the proposal background is modified to

$\delta$61

and $\delta$62 is tuned so that $\delta$63 dominates the unknown background over a localized signal region $\delta$64 (Banerjee et al., 19 May 2026).

This is where the paper uses the notion of a relaxed compensator most explicitly: increasing $\delta$65 makes $\delta$66 more negative and inference more conservative, whereas decreasing $\delta$67 moves $\delta$68 up toward $\delta$69 from below, thereby “relaxing” conservativeness while preserving validity. The reported case study illustrates this trade-off: $\delta$70 yielded $\delta$71 with $\delta$72; $\delta$73 gave $\delta$74 with $\delta$75; and $\delta$76 made detection fail with $\delta$77 (Banerjee et al., 19 May 2026).

This statistical usage is conceptually close to the stochastic-process usages in one respect: the compensator isolates the part of uncertainty that matters for the inferential target. But the object is no longer a predictable process. It is a one-dimensional geometric coefficient governing conservativeness in both the proposed framework and misspecified likelihood-ratio procedures.