Design of Input-Output Observers for a Population of Systems with Bounded Frequency-Domain Variation using $DK$-iteration

James Richard Forbes; Timothy Everett Adams

REVIEW 2 major objections 5 minor 15 references

Design of Input-Output Observers for a Population of Systems with Bounded Frequency-Domain Variation using $DK$-iteration

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single robust correction filter gives guaranteed state-estimation accuracy across a whole population of similar systems.

desk verdict Solid practical contribution — the filter design is sensible and the experiment is clean — but the paper overclaims the guarantee by certifying mu on a 61-point grid, and the validation is in-sample. read the letter →

arxiv 2509.07201 v2 pith:C334HC7U submitted 2025-09-08 eess.SY cs.SY

Timothy Everett Adams , James Richard Forbes This is my paper

classification eess.SYcs.SY

keywords stateobserverrobustcontrolDK-iterationstructuredsingularvalueuncertaintymodelingsystemidentificationpopulationvariationflexiblejointmanipulator

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Observers for batches of similar devices usually need either a gain tuned per device or a single filter designed from one nominal model with no accuracy guarantee. This paper proposes a middle path: identify a model for each device, characterize the population's variation in the frequency domain with an uncertainty weight, and synthesize one correction filter by solving a robust performance problem with DK-iteration. The result is a guaranteed worst-case estimation performance for every system that lies inside the variation description, and the same filter can be paired with any future device model that also lies inside it. The method is demonstrated on a flexible-joint robotic manipulator with four different joint stiffness configurations, where the single filter's absolute estimation errors are comparable to a separately tuned Kalman filter for each configuration.

What carries the argument

The central mechanism is the inverse multiplicative input uncertainty model G(s) = G0(s)(1 - W_Delta(s)Delta(s))^{-1} together with the structured singular value (SSV) robust performance condition mu_hatDelta(N(jw)) < 1 evaluated on the observer error dynamics. The uncertainty weight W_Delta(s) converts measured population spread into a norm-bounded perturbation set; the performance weights W_d, W_n, W_e, and W_nu convert 'good estimation' into frequency-shaped signals; and DK-iteration alternates H-infinity synthesis of the correction filter K(s) with D-scaling analysis to drive the SSV upper bound below 1. The D(s) scaling fits a stable minimum-phase transfer function to frequency-wise sca

What would settle it

Take the filter synthesized in the paper and pair it with a device whose measured frequency response violates the fitted overbound max_i |E_iI(jw)| <= |W_Delta(jw)| at any frequency, or run the same robustness analysis on a much denser frequency grid than the 61 points used; finding a frequency where mu_hatDelta(N(jw)) > 1, or experimentally observing estimation error above the guaranteed level for an in-set device, would falsify the claim that a single correction filter certifies performance across the population.

Watch

Extended reading notes

Core claim

The paper's claim is that population-level observer design can be posed and solved as a frequency-domain robust performance problem, so that one correction filter serves all devices. Each device keeps its own identified LTI model, while the population spread is captured by an inverse multiplicative input uncertainty description G(s) = G0(s)(1 - W_Delta(s)Delta(s))^{-1}, with W_Delta a fitted stable weight that overbounds the measured residual between each device and the nominal model. The observer error dynamics are augmented with performance weights and the structured singular value condition mu_hatDelta(N(jw)) < 1 is enforced on a frequency grid via DK-iteration. When the condition holds,

Load-bearing premise

The entire guarantee rests on the fitted uncertainty weight W_Delta(s) actually overbounding every device's true deviation from the nominal model, including behavior outside the fitted frequency range and errors from the identification step; if any real device escapes that bound, the certified performance guarantee does not apply to it.

Editorial extensions

If this is right

Only one correction filter needs to be synthesized for an entire batch; per-device models can be obtained by a standard identification procedure and paired with the same filter.
A new device can be added to the observer population after the design is complete, without re-synthesis, provided its identified model satisfies the same frequency-domain variation bound.
The robust performance certificate replaces ad hoc confidence in a nominal-model filter with a formal worst-case accuracy claim for the whole variation set.
In the flexible-joint experiments, the single filter's error distributions across four stiffness configurations are comparable to a tailored Kalman filter, with differences on the order of a few tenths of a degree.
The design flow is not tied to the particular manipulator; any population whose frequency-response residuals can be represented by the inverse multiplicative uncertainty model could use the same procedure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The guarantee is only as broad as the fitted weight W_Delta: a device whose true behavior escapes the bound, through unmodeled nonlinearity or behavior outside the measured frequency range, falls outside the certified population even if the observer still happens to work in practice.
Because the SSV condition is checked on a 61-point grid from 0.01 to 25 Hz, the all-frequency guarantee implicitly assumes grid verification is sufficient; a denser or wider grid would be a direct test of that assumption.
A natural validation experiment is to take a new joint-stiffness configuration not used in the identification, confirm it lies inside the fitted W_Delta bound, and check that the observed estimation error stays below the guaranteed level.
The conservatism of the single-filter design could be reduced by fitting a tighter W_Delta or allowing higher-order D(s) scalings, at the cost of more complex synthesis, so the method has a tunable trade-off between filter order and population coverage.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Tim, this one is worth a look if you care about observer design for populations. The idea: instead of designing a separate correction filter for each identified model, characterize the population spread as a multiplicative uncertainty set around a nominal model, then run DK-iteration on the input-output observer error dynamics to get one robust correction filter. The individual ingredients are all standard—input-output observers, multiplicative uncertainty, DK-iteration—but the packaged procedure is genuinely new, and the experimental demonstration on a flexible-joint arm with four stiffness configurations is clear. They report error comparable to a tailored steady-state Kalman filter, which is a strong result given the filter is not tuned to any individual.

The core derivation is standard and the robust-performance argument is valid conditional on the uncertainty model and the mu condition. The code is posted, which helps. My main reservation is the verification of the central guarantee. Equation (6) requires mu-hat < 1 for all frequencies, but the numerical evidence is a 61-point log grid from 0.01 to 25 Hz. No density argument or all-frequency H-infinity bound on the scaled interconnection is given, and the tail above 25 Hz is not trivial because of the high-frequency penalty Wnu. This might be fixable by computing the H-infinity norm of D N D^{-1} after the final D(s) fit, which is standard practice in mu-synthesis, but the paper doesn't report that. Right now the claim 'compatible with all systems in the set' rests on a sampled frequency axis.

The second issue is that the uncertainty weight W_delta is fitted to the same four identified models used in the demonstration. So the experimental section is an in-sample illustration, not a predictive test. That doesn't break the method—if the weight legitimately overbounds the population, the guarantee still holds—but it does mean the validation is not independent. A baseline with a single nominal filter would also have been nice, to show what the robust filter adds over a naive design. The comparison to tailored KFs is good but not sufficient to demonstrate the need.

Overall: a reasonable engineering contribution, clearly written, with a real experiment. It deserves review, but the referee should push on the frequency-grid certification and ask for a more rigorous all-frequency check. I would cite it if I were working on robust observers; it's a clean building block.

Referee Report

2 major / 5 minor

Summary. The paper proposes a methodology for designing a single dynamic correction filter for input-output observers used across a population of LTI systems with bounded frequency-domain variation. The design pipeline is: (1) identify individual device models, (2) characterize population variation by an inverse multiplicative input uncertainty weight W_Delta(s), (3) construct a generalized plant from observer error dynamics with frequency-dependent performance weights, (4) synthesize one robust correction filter K(s) by DK-iteration, and (5) pair this filter with each device's identified model to form the observer. The central claim is that if the structured singular value condition mu_hatDelta(N(jw)) < 1 in Eq. (6) is achieved for all frequencies, then the observer delivers a guaranteed level of performance for every model in the variation set. The method is demonstrated experimentally on a two-degree-of-freedom flexible-joint manipulator with four joint-stiffness configurations, where the robust observer is compared with a per-configuration steady-state Kalman filter.

Significance. If the guarantee can be made rigorous, the method is practically valuable: it reduces per-device observer tuning to a single filter synthesis and provides a certificate for all systems consistent with the uncertainty characterization. The theoretical reduction from observer error dynamics to a robust performance problem is standard and the derivation of N(s) = F_l(P(s), K(s)) and the use of Eq. (6) is internally consistent. The authors provide reproducible code, real-hardware experiments, and a fair comparison against a tailored Kalman filter. The main gaps are the numerical certification of the all-frequency condition and the empirical basis for treating the fitted W_Delta as a valid overbound of the population variation.

major comments (2)

[§IV, Fig. 10 and Eq. (6)] The central guarantee is Eq. (6), stated as mu_hatDelta(N(jw)) < 1 for all w in [0, infinity). The only numerical support is the finite grid used in Section IV: 61 logarithmically spaced points from 0.01 Hz to 25 Hz, after which the authors conclude from Fig. 10 that the robust performance criterion is achieved. A finite grid cannot certify an all-frequency condition for rational LTI interconnections; narrow SSV peaks can occur between grid points. Moreover, Fig. 9 shows the high-pass performance weight W_nu is still active at and beyond 25 Hz, so the omitted tail cannot be dismissed by inspection. To support the stated guarantee, the authors should either verify Eq. (6) at all frequencies via a state-space/KYP or LMI test on the mu upper bound, use an adaptive-grid or branch-and-bound certification, or explicitly weaken the claim to certification on the tested grid. As written, the clai
[§IV, Fig. 7, Eq. (14)] The uncertainty weight W_Delta(s) is fitted to max(E_iI)(jw), where E_iI is computed from the same four identified models that are later used in the experimental demonstration. This makes the demonstration a consistency check rather than an independent predictive test of the population claim. More importantly, the robust performance guarantee applies only if W_Delta is a valid overbound for all devices in the population, including unmodeled nonlinearities, identification errors, and frequencies outside the fitted range. The paper does not provide evidence that W_Delta bounds variation beyond the four fitted models. I recommend either adding a leave-one-out or withheld-device validation, or explicitly stating that the guarantee is conditional on W_Delta being a valid overbound and providing supporting evidence. Without this, the conclusion that the procedure is effective for the populatio

minor comments (5)

[Title and throughout] 'usingDK-iteration' and similar instances should be 'using DK-iteration' (missing space).
[§II.B, Step 3] The text says D(s) is fit as a 'stable and non-minimum phase transfer matrix.' In standard DK-iteration, D(s) should be stable and minimum-phase. The W_Delta weight in Section IV may be non-minimum phase, but the D-scaling should be minimum-phase. Please clarify or correct the terminology.
[Eq. (8)] The notation 'D(jw)-1' would be clearer as 'D(jw)^{-1}' to avoid confusion with subtraction.
[Fig. 10] The legend labels the curves 'Iteration 0' and 'Iteration 1'; please clarify whether iteration 0 is the initial H-infinity design and iteration 1 is the first DK pass, and state the order of the D(s) fit used in each.
[Fig. 12] The box plots aggregate data across the four stiffness configurations. Please state the number of samples per box and whether configurations are pooled, so the comparison with the Kalman filter can be interpreted quantitatively.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the robust-performance guarantee follows from the standard SSV theorem; self-citations are not load-bearing.

full rationale

The paper's central claim is conditional and rests on standard robust control theory, not on its own fitted quantities. The synthesis procedure uses DK-iteration to make the structured singular value upper bound (8) less than 1 on a frequency grid, and then invokes the standard robust performance theorem (Eq. (6), from Skogestad and Postlethwaite [1]) to conclude that the resulting filter guarantees the weighted performance objective for every perturbation Delta with ||Delta||_infinity <= 1. This is an external theorem, not a restatement of the fitted uncertainty weight. The weight W_Delta is fitted to the residuals of the same four identified models used in the experiment, so the experimental demonstration is not an out-of-sample predictive test; however, the paper does not claim to predict performance on unseen devices. Instead, it demonstrates that a single correction filter performs comparably to a tailored Kalman filter on the population used for characterization. That is an experimental comparison, not a derivation that reduces to its inputs. The finite-frequency grid (61 points from 0.01 Hz to 25 Hz) used to verify condition (6) is a correctness limitation: the paper asserts the all-frequency condition based on a sampled grid without a density argument or high-frequency analysis. But this is an evidential gap, not circularity by construction. The self-citations ([11] for LMI background and [16] for the dkpy software) are not load-bearing for the central robust-control argument; the core mu-analysis and DK-iteration theory is cited to an external source [1]. Therefore no specific circular step can be exhibited, and the appropriate circularity score is 0.

Assumptions & free parameters 4 free parameters · 7 assumptions · 0 invented entities

The central claim rests on a fitted uncertainty weight, hand-chosen performance weights, and the implicit assumption that a finite frequency grid certifies all-frequency robust performance. These are the main sources of potential fragility, along with the linear regime assumption.

free parameters (4)

Uncertainty weight W_Delta(s) = Stable non-minimum phase transfer function overbounding max residuals (not explicitly given)
Fitted to overbound the maximum residuals EiI(jw) across the four identified models. The robust performance guarantee depends on this weight; if it underestimates the actual variation, the guarantee is void.
Performance weights Wd(s), Wn(s), We(s), Wnu(s) = Magnitude responses shown in Fig. 9
Chosen by the designer to encode disturbance/noise/error/correction frequency content. The 'given level of estimation performance' is defined by these weights, so the guarantee is relative to this choice.
Order of D(s) fit = 8th order
Synthesis hyperparameter in DK-iteration. The order affects the scaling matrix fit and the conservatism of the mu upper bound.
Frequency grid for SSV evaluation = 61 points from 0.01 Hz to 25 Hz
The robust performance condition is verified on this grid. The guarantee implicitly assumes the grid is dense enough to represent all-frequency behavior.

assumptions (7)

domain assumption The devices are LTI and operate in a linear small-perturbation regime around an equilibrium.
Stated in Section IV: 'the estimation is performed in a linear regime of small perturbations about the equilibrium point in which both links are colinear and stationary.' The guarantee applies only to the linearized dynamics.
domain assumption The measurement model is a static mapping with no feedthrough.
Assumed in Section II-A: 'It is assumed that the measurement model is a static mapping with no feedthrough.'
standard math The perturbation Delta(s) is any LTI system with ||Delta||_inf <= 1.
Standard robust control assumption from Skogestad, used in Eq. 4 to define the uncertainty set.
ad hoc to paper The inverse multiplicative input uncertainty model with weight W_Delta overbounds all population variation.
The weight is fitted to the residuals of the identified models. It is not proven to cover all possible devices in the population; this is the weakest assumption.
ad hoc to paper The performance weights Wd, Wn, We, Wnu encode the desired estimation performance specifications.
These weights are design choices that define the normalized signals w and z. If they do not reflect the actual performance metric of interest, the guarantee is not meaningful.
standard math DK-iteration converges to a controller achieving the robust performance criterion.
DK-iteration is a standard heuristic that works well in practice but does not guarantee global optimality. The paper relies on this algorithm.
ad hoc to paper The finite frequency grid is sufficient to certify mu < 1 over all frequencies.
The SSV is evaluated at 61 discrete frequencies. The paper does not provide a worst-case or analytic guarantee for frequencies between grid points.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Design of Input-Output Observers for a Population of Systems with Bounded Frequency-Domain Variation using $DK$-iteration." pith.science (2026). https://pith.science/paper/C334HC7U

@misc{pith2026250907201,
  author       = {Pith},
  title        = {Pith review of: Design of Input-Output Observers for a Population of Systems with Bounded Frequency-Domain Variation using $DK$-iteration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C334HC7U}},
  note         = {Machine review of arXiv:2509.07201}
}

abstract

This paper proposes a linear input-output observer design methodology for a population of systems in which each observer uses knowledge of the linear time-invariant dynamics of the particular device. Observers are typically composed of a known model of the system and a correction mechanism to produce an estimate of the state. The proposed design procedure characterizes the variation within the population in the frequency domain and synthesizes a single robust correction filter. The correction filter is compatible with all system models that satisfy the variation characterization such that a given level of estimation performance is guaranteed. This is accomplished by posing a robust performance problem using the observer error dynamics and solving it using $DK$-iteration. The design procedure is experimentally demonstrated on a flexible joint robotic manipulator with varied joint stiffnesses. It is shown that the proposed method that uses a single correction filter achieves comparable estimation performance to a method that uses a correction gain tailored toward each joint stiffness configuration.

Figures

Figures reproduced from arXiv: 2509.07201 by the authors.

**Figure 2.** Input-output observer. K(s) G(s) C −n(s) du(s) dx(s) ρ(s) ν(s) ey(s) ex(s) + − + + + + [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Input-output observer error dynamics. where C ∈ R ny×nx is a constant matrix that maps the states x(s) to the outputs y(s). It is assumed that the measurement model is a static mapping with no feedthrough. An attractive observer structure is the input-output observer, depicted in [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: General robust control problem block diagram. [PITH_FULL_IMAGE:figures/full_fig_p003_4.png]

**Figure 5.** Figure 5: Picture of Quanser 2 DOF Serial Flexible Joint with colorized rigid bodies (blue: shoulder motor hub; orange: shoulder link; red: elbow motor hub; green: elbow link). TABLE I: Flexible joint manipulator system variables. Parameter Unit Symbol Shoulder motor current A i…

**Figure 6.** Figure 6: Magnitude response of the identified process models [PITH_FULL_IMAGE:figures/full_fig_p004_6.png]

**Figure 7.** Figure 7: Magnitude response of the inverse multiplicative input [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]

**Figure 8.** Figure 8: Robust observer generalized plant diagram. [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 11.** Figure 11: Estimated angular positions for the robust state [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

**Figure 12.** Figure 12: Box plot of the absolute error of the estimated [PITH_FULL_IMAGE:figures/full_fig_p006_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 13 canonical work pages

[1]

Skogestad and I

S. Skogestad and I. Postlethwaite,Multivariable Feedback Con- trol: Analysis and Design, 2 edition. Chichester: Wiley-Interscience, Nov. 4, 2005, 592 pp

work page 2005
[2]

Williams and D

R. Williams and D. Lawrence,Linear State-Space Control Systems, 1st ed. John Wiley & Sons, Ltd, 2007

work page 2007
[3]

Fault detection methods: A literature survey,

D. Miljkovi ´c, “Fault detection methods: A literature survey,” in 2011 Proceedings of the 34th International Convention MIPRO, May 2011, pp. 750–755

work page 2011
[4]

Power System Real-Time Monitoring by Using PMU-Based Robust State Estimation Method,

J. Zhao et al., “Power System Real-Time Monitoring by Using PMU-Based Robust State Estimation Method,”IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 300–309, Jan. 2016

work page 2016
[5]

B. D. O. Anderson and J. B. Moore,Optimal Filtering. Prentice-Hall Information and System Sciences Series, 1979, 370 pp

work page 1979
[6]

Simon,Optimal State Estimation: Kalman, H [Infinity] and Non- linear Approaches

D. Simon,Optimal State Estimation: Kalman, H [Infinity] and Non- linear Approaches. Hoboken, N.J: Wiley-Interscience, 2006, 526 pp

work page 2006
[7]

Green and D

M. Green and D. J. N. Limebeer,Linear Robust Control. Courier Corporation, Sep. 19, 2012, 562 pp

work page 2012
[8]

A frequency domain approach to state estimation,

H. J. Marquez, “A frequency domain approach to state estimation,” Journal of the Franklin Institute, vol. 340, no. 2, pp. 147–157, Mar. 1, 2003

work page 2003

Show all 15 references

[9]

Robust state observer design with application to an industrial boiler system,

H. J. Marquez and M. Riaz, “Robust state observer design with application to an industrial boiler system,”Control Engineering Practice, vol. 13, no. 6, pp. 713–728, Jun. 1, 2005

2005
[10]

Ljung,System Identification: Theory for the User

L. Ljung,System Identification: Theory for the User. Prentice Hall, 1999

1999
[11]

LMI Properties and Applications in Systems, Stability, and Control Theory

R. J. Caverly and J. R. Forbes. “LMI Properties and Applications in Systems, Stability, and Control Theory.” arXiv:1903.08599, pre-published

1903 arXiv
[12]

2 DOF Serial Flexible Joint,

Quanser. “2 DOF Serial Flexible Joint,” Accessed: Jan. 4, 2025. [On- line]. Available:https://www.quanser.com/products/2- dof-serial-flexible-joint/

2025
[13]

Pintelon and J

R. Pintelon and J. Schoukens,System Identification: A Frequency Domain Approach. John Wiley & Sons, Inc., 2012. [14]System Identification Toolbox, MATLAB. Accessed: Aug. 12, 2025. [Online]. Available:https : / / www . mathworks . com / products/sysid.html

2012
[15]

FIR Filter Design via Spectral Factorization and Convex Optimization,

S.-P. Wu, S. Boyd, and L. Vandenberghe, “FIR Filter Design via Spectral Factorization and Convex Optimization,” inApplied and Computational Control, Signals, and Circuits, B. N. Datta, Ed., Boston, MA: Birkh ¨auser Boston, 1999, pp. 215–245

1999
[16]

Dahdah and J

S. Dahdah and J. R. Forbes,Decargroup/dkpy, version v0.1.8, Zenodo, 2024. [Online]. Available:https : / / github . com / decargroup/dkpy

2024

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Skogestad and I

S. Skogestad and I. Postlethwaite,Multivariable Feedback Con- trol: Analysis and Design, 2 edition. Chichester: Wiley-Interscience, Nov. 4, 2005, 592 pp

work page 2005

[2] [2]

Williams and D

R. Williams and D. Lawrence,Linear State-Space Control Systems, 1st ed. John Wiley & Sons, Ltd, 2007

work page 2007

[3] [3]

Fault detection methods: A literature survey,

D. Miljkovi ´c, “Fault detection methods: A literature survey,” in 2011 Proceedings of the 34th International Convention MIPRO, May 2011, pp. 750–755

work page 2011

[4] [4]

Power System Real-Time Monitoring by Using PMU-Based Robust State Estimation Method,

J. Zhao et al., “Power System Real-Time Monitoring by Using PMU-Based Robust State Estimation Method,”IEEE Transactions on Smart Grid, vol. 7, no. 1, pp. 300–309, Jan. 2016

work page 2016

[5] [5]

B. D. O. Anderson and J. B. Moore,Optimal Filtering. Prentice-Hall Information and System Sciences Series, 1979, 370 pp

work page 1979

[6] [6]

Simon,Optimal State Estimation: Kalman, H [Infinity] and Non- linear Approaches

D. Simon,Optimal State Estimation: Kalman, H [Infinity] and Non- linear Approaches. Hoboken, N.J: Wiley-Interscience, 2006, 526 pp

work page 2006

[7] [7]

Green and D

M. Green and D. J. N. Limebeer,Linear Robust Control. Courier Corporation, Sep. 19, 2012, 562 pp

work page 2012

[8] [8]

A frequency domain approach to state estimation,

H. J. Marquez, “A frequency domain approach to state estimation,” Journal of the Franklin Institute, vol. 340, no. 2, pp. 147–157, Mar. 1, 2003

work page 2003