Control and Identification in Least-Squares Self-Tuning Regulation with Unknown Input Gain
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.
Ada: I'm Ada.
Emmy: And I'm Emmy, here with Kolya.
Ada: We're looking at the paper: "Control and Identification in Least-Squares Self-Tuning Regulation with Unknown Input Gain".
Paper summary: Ada: So we're digging into how this paper proves that strong consistency and optimal tracking are achievable for this self-tuning regulator with unknown input gain. We just touched on the high-level claim, so let's look closer at the structure they built to get there.
Emmy: They focus on a single input single output system where the plant polynomials are defined by some formula, and their objective is keeping both the average input and output energy bounded while making that average squared tracking error converge to the noise variance.
Kolya: That sounds like a very specific target. What kind of systems are they talking about when they define those plant polynomials? Are we talking about simple linear systems, or something more complex?
Ada: They're looking at minimum-phase linear systems, and the paper establishes this joint guarantee for ordinary least squares with certainty equivalent control without reference excitation, gain adjustment or added probing signals >
Emmy: The core of their consistency result hinges on constructing an auxiliary least-squares prediction driven by process noise whose cumulative squared difference from a bounded reference signal has a logarithmic lower bound.
Kolya: A logarithmic lower bound sounds like it's giving us some kind of information about how the noise behaves over time, which is useful since we don't know the exact gain >
Ada: Precisely. The authors argue that if that gain error were to persist, this noise information and the actual least-squares recursion would create incompatible lower and upper bounds on the same weighted squared sum >
Emmy: That contradiction is how they establish gain consistency before they even prove stability for the overall regulator >
Kolya: So, it seems like they are using this noise information as a kind of internal compass to check if their estimates are actually converging correctly, even when things look okay on the surface >
Conclusion: Ada: Looking at the whole paper, "Control and Identification in Least-Squares Self-Tuning Regulation with Unknown Input Gain," it's about showing that this self-tuning regulator can perform well without knowing the input gain.
Emmy: They achieve strong consistency when at least two parameters are estimated, meaning they get their true parameter values almost surely >
Kolya: So what does that mean practically? If we have a system with two unknown parameters, like a mass and a damping coefficient, this method should eventually give us those correct values without us having to add extra stuff to the control loop >
Ada: Right. The title itself points to the main struggle: controlling and identifying when the input gain is unknown. The authors show that you can do both things simultaneously in this setup >
Emmy: What I find most interesting is how they use those LS identities and an auxiliary prediction driven by noise to constrain the estimation process, showing that bounded deterministic reference signals can't cancel out the growth of that noise prediction >
Kolya: That part about the contradiction establishing consistency before stability is what really stands out to me. It’s a clever way to prove something about convergence using only what we can measure from the noise itself >
Ada: So, it’s a result that suggests for certain types of systems, you can build a robust self-tuning system that handles the uncertainty of the input gain quite well without needing those external probes or adjustments usually required >
ZHAOBO LIU
math.OC
Submitted: 2026-09-10
Updated: 2026-11-17
Comments: 22 pages, 1 figure
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 80/100
The gist: Strong consistency and optimal tracking of the Åström–Wittenmark self-tuning regulator with unknown input gain is established for a class of minimum-phase linear systems with martingale…
Key concepts
- Self-Tuning Regulator (STR)
- This controller combines recursive least squares (RLS) for parameter estimation with certainty-equivalent control. It uses current parameter estimates to calculate the necessary control action based on a known reference signal, aiming to keep system performance optimal.
- Gain Consistency Mechanism
- This mechanism proves that the unknown input gain can be consistently estimated even if it is unknown. It relies on comparing bounds derived from two different estimation methods: one using process noise and another using the actual system recursion. A contradiction forces the gain error to zero.
- Auxiliary Least-Squares Prediction
- This is a secondary least-squares calculation specifically designed to capture information from the process noise. The paper shows that this auxiliary prediction has a logarithmic lower bound related to noise, which is crucial for constraining the parameter estimation errors in the main system.
Terminology
Summary
Strong consistency and optimal tracking of the Åström–Wittenmark self-tuning regulator with unknown input gain is established for a class of minimum-phase linear systems with martingale difference noise The key is a logarithmic lower bound on the cumulative squared difference between an auxiliary least-squares prediction of the noise and the reference If the gain error persisted, this noise information and the actual least-squares recursion would give incompatible lower and upper bounds on the same weighted squared sum. p. 0
Problem Formulation
The study focuses on a single-input single-output system where plant polynomials are defined by (1.2) The objective is to keep the average input and output energy bounded and make the average squared tracking error converge to the noise variance. p. 0 The self-tuning regulator (STR) combines ordinary unweighted least squares (LS), implemented recursively, with certainty-equivalent (CE) control. p. 0 The controller divides by its estimate, even when an exact zero divisor is avoided. p. 1
Main Results
Under Assumptions 2.1 and 2.2 and the condition (2. p. 110), the recursion satisfies almost sure stability and optimal tracking for each bounded reference. p. 1 When at least two parameters are estimated, the entire parameter estimate converges almost surely to its true value. p. 1 The result holds for the original LS-based STR without added probing. p. 1
Gain Consistency Mechanism
The consistency result hinges on constructing an auxiliary LS prediction driven by process noise whose cumulative squared difference from a bounded reference has a logarithmic lower bound. This information is transferred to the actual recursion using Guo’s estimates along subsequences on which the gain error stays bounded away from zero. p. 1 If the gain failed to converge, these bounds and the noise information would impose incompatible upper and lower bounds on the same weighted energy sum, establishing gain consistency before stability. p. 1
Information Constraints
The analysis involves several key mathematical tools to constrain estimation. For example, Lemma 3. p. 42 shows that a bounded deterministic reference cannot cancel the growth of the auxiliary LS prediction. p. 6 Furthermore, Lemma D. p. 53 demonstrates that the term n(Jn + Hn)mnµn is o(1), where Jn and Hn are related to information matrices. p. 2 This implies that the LS identities force Jn = o(µn). p. 9
Conclusion
The ordinary LS-based STR with unknown input gain achieves almost sure stability and optimal tracking for each bounded reference, and strong consistency when d ≥ 2. The key observation is that persistent gain error would force incompatible bounds on the same weighted squared sum. p. 0 An auxiliary LS prediction extracts information from process noise that a bounded reference cannot cancel, whereas the identities of the actual LS recursion impose a smaller upper bound. p. 1 p. 1 This contradiction establishes gain consistency before stability is known. p. 0 The asymptotic properties do not ensure finite expected control energy over a finite horizon, even with bounded continuous noise. p. 1
References
[2] Bernard Bercu. Weighted estimation and tracking for ARMAX models [3] Bernard Bercu and Victor Vazquez. p. 1 A new concept of strong controllability via the Schur complement for ARX models in adaptive tracking [4] Bernard Bercu and Victor Vazquez. p. 0 On the usefulness of persistent excitation in ARX adaptive tracking [5] Han-Fu Chen. p. 0 Recursive system identification and adaptive control by use of the modified least squares algorithm [6] Han-Fu Chen and Lei Guo. p. 0 Convergence rate of least-squares identification and adaptive control for stochastic systems [7] Han-Fu Chen and Lei Guo. p. 0 Asymptotically optimal adaptive control with consistent parameter estimates [8] V. p. 0 Ž. Filipović. p. 1 Convergence and optimality of stochastic adaptive control scheme when the disturbance is non-Gaussian [9] Graham C. p. 1 Goodwin, Peter J. Ramadge, and Peter E. Caines. Discrete time stochastic adaptive control [10] Lei Guo. p. 0 Further results on least squares based adaptive minimum variance control [11] Lei Guo. p. 1 Convergence and logarithm laws of self-tuning regulators [12] Lei Guo and Han-Fu Chen. p. 0 The Åström–Wittenmark self-tuning regulator revisited and ELS-based adaptive trackers [13] P. p. 0 R. Kumar. Convergence of adaptive control schemes using least-squares parameter estimates [14] P. p. 0 R. Kumar and Laurent Praly. p. 1 Self-tuning trackers [15] Tze Leung Lai and Ching Zong Wei. p. 0 Least squares estimates in stochastic regression models with applications to identification and control of dynamic systems [16] Tze Leung Lai and Ching Zong Wei. p. 0 On the concept of excitation in least squares identification and adaptive control [17] Tze Leung Lai and Ching Zong Wei. p. 0 Asymptotically efficient self-tuning regulators [20] S. p. 0 P. Meyn and P. E. Caines. The zero divisor problem of multivariable stochastic adaptive control [24] Yaqi Zhang and Lei Guo. p. 0 Convergence of self-tuning regulators under conditional heteroscedastic noises with unknown high-frequency gain [25] Institute for Advanced Study, Shenzhen University, Shenzhen 518060, China Email address: liuzhaobo@szu. p. 0edu.cn (Note: The provided references list contains IDs that are not explicitly mapped to the arXiv ID 2609. p. 011831 in the text, so I have used the provided reference numbers as placeholders for the required token format. p. 0)
Pruning redundant reasoning tokens
The paper investigates strong consistency and optimal tracking of the Åström–Wittenmark self-tuning regulator with unknown input gain Existing results establish stability, optimal tracking and parameter consistency for the unmodified recursion under growth conditions on reference information Other approaches obtain performance guarantees by adjusting the estimates used in feedback or adding decaying probing signals. For a class of minimum-phase linear systems with martingale difference noise, joint guarantees are established for ordinary least squares with certainty-equivalent control without reference excitation, gain adjustment or added probing. p. 0
How the training signal is built
The controller uses the current estimate and the known reference y∗n+1 to select un. The parameter estimate is updated by the ordinary recursive least squares (RLS) algorithm. p. 2 The CE control equation is given by (2. p. 03).
Where the gains come from
The convergence of the gain estimate is proven by showing that persistent gain error would force incompatible bounds on the same weighted squared sum. An auxiliary LS prediction extracts information from process noise that a bounded reference cannot cancel, whereas the identities of the actual LS recursion impose a smaller upper bound. This contradiction establishes gain consistency before stability. p. 0 The final strong consistency for d ≥ 2 is achieved when the estimated gain is eventually bounded away from zero. p. 1
Noise Information and LS Identities
The analysis relies on LS identities and auxiliary prediction to constrain estimation in the direction of the common parameter scale. Lemma 3.2 shows that a bounded deterministic reference cannot cancel this growth. p. 6 The information update and the control equation give ∆Jn = Jn+1 − Jn = (πn + k T n xn) 2/ (1 + λn). p. 6
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Strong consistency for parameter estimates is achieved when at least two parameters are estimated, which allows for identification of all unknown coefficients in minimum variance tracking without reference excitation or added probing signals.
-
The system can achieve almost sure stability and optimal tracking for each bounded reference signal even with an unknown input gain, provided that the degree of the plant polynomials meets certain conditions (Assumption 2.2).
-
The improved AI system can maintain the average squared tracking error at the noise variance, which is the minimum attainable with known plant coefficients, almost surely.
-
For systems where at least two parameters are estimated, all parameter estimates converge almost surely to their true values, ensuring full identification of the unknown plant coefficients without needing reference excitation.
Related papers
- An Accelerated Proximal Gradient Method for Bures-Wasserstein Space
- FOMTools: Development and AI-Assisted Code Generation for First-Order Methods
- Lions and Muons: Optimization via Stochastic Frank-Wolfe under Heavy-Tailed Noise
- Dynamic Service Recommendation with Congestion-Dependent Joining: Near-Optimal and Constant-Factor Approximation Algorithms
- Integrating Learning and Optimization for Open Transport Capacity Pooling under Disruption
- Zeroth-Order Gradient Estimators under Decision-Dependent Noise