Recursive Least Squares Filter¶
A sequential linear-model estimator that maintains least-squares matrix state and uses each new prediction error to update its coefficients.
Core Idea¶
A recursive least squares (RLS) filter adapts the coefficients of a linear predictor as input–response pairs arrive. Instead of solving a new batch least-squares problem after every sample, it maintains a coefficient estimate and matrix information representing the prior data. For the next pair, it computes a prediction, forms the a priori error, calculates a gain shaped by that matrix and the current input, and updates both the coefficients and matrix state. Conventional RLS algebra follows a cumulative squared-error objective with equal or exponentially decaying historical weights, subject to the stated initialization and arithmetic assumptions.[1][2]
The identity is the recurring linear residual → data-adaptive gain → coefficient/matrix correction relation. A forgetting factor below one is useful for changing systems, but one is a valid no-forgetting case. A particular inverse-matrix implementation, guaranteed superiority to LMS, or exactness in finite-precision arithmetic is not part of the definition.[2][3]
Structural Signature¶
Sig role-phrases:
- Sequential linear prediction problem. Each new regressor \(u_t\) and desired response \(d_t\) supplies a residual against a coefficient vector \(w\). Without a linear-model squared-error objective, a similar-looking gain recursion need not be RLS.[1]
- Maintained least-squares state. The algorithm carries \(w\) and inverse-correlation-like information, or an algebraically equivalent square-root/QR factorization, summarizing earlier weighted observations. Re-solving the complete batch from stored observations is least squares but not the recursive algorithm.[1][2]
- Error-gain correction. The new prior error \(e_t=d_t-w_{t-1}^{T}u_t\) is multiplied by a gain depending on the current input and maintained matrix; the matrix is also updated. A fixed scalar-step gradient rule without that quadratic-fit state is LMS-like rather than RLS.[1][2]
The weighting convention, initialization, matrix dimensions and numerical realization must be named when comparing implementations. Exponential discounting and faster-than-LMS convergence are contingent settings or performance properties, not additional necessary roles.[2][3]
What It Is Not¶
- Not batch linear least squares itself. A batch solver minimizes residuals over a dataset, whereas RLS specifies a sequential update of the solution's sufficient matrix state.[1]
- Not generic recursion. Reusing yesterday's estimate alone is insufficient; the specific least-squares matrix/gain relation matters.
- Not LMS. LMS uses a gradient-style correction rather than the RLS data-adaptive inverse-correlation gain.[2]
- Not necessarily Bayesian filtering. Some Kalman formulations have related algebra under specified models, but a posterior prediction–likelihood cycle is not required for the RLS identity.[2]
- Not necessarily forgetting. Setting \(\lambda=1\) retains equal historical weighting, while \(0<\lambda<1\) discounts old samples.[2]
Scope of Application¶
In adaptive FIR system identification, a window of recent input samples predicts an unknown filter's output. The estimated tap coefficients and matrix state update from each desired-output error. MathWorks' worked example varies an FIR filter and the RLS forgetting factor, making tracking behavior visible in a controlled simulation; it also warns that a too-low forgetting factor can prevent convergence in that setup.[3]
In online parameter estimation, the same form can update the above-dead-band line-fit segment of a valve model, \(y(t)=k(t)u(t)+b(t)+e(t)\): the regressor \([u(t),1]\) predicts output and the two coefficients are gain and offset. The source's complete valve response is piecewise, so the line is not asserted outside that active operating range. This is not a literal audio-filter use, but it fills the same sequential residual, matrix information and gain-correction roles.[4][2]
Clarity¶
The phrase “exact least squares” needs qualification. In exact algebra, conventional recursions can maintain the minimizer of a declared weighted or initialized quadratic objective. With finite arithmetic, ill-conditioned excitation, inappropriate initialization or unstable matrix updates, a computed result may drift. MathWorks documents square-root variants to preserve positive-definite state; that numerical safeguard is an implementation issue, not a new identity.[1][2]
“Faster convergence” is similarly conditional. The comparison depends on excitation, filter size, tuning and the performance criterion. Official documentation says RLS/forgetting-factor forms typically have better convergence properties than gradient methods, while its own FIR example shows failure when forgetting is tuned too aggressively. Neither statement is a universal guarantee.[2][3]
Manages Complexity¶
The signature gives a compact way to audit algorithms called adaptive filters. First locate the linear prediction and residual. Next ask what persistent matrix or factor state encodes prior observations. Then derive the gain and the coupled coefficient/state update. If the matrix state is absent, a mere gain-looking formula may be LMS; if all data are refitted each time, the calculation is batch least squares; if a transition prior and likelihood define a posterior, it may be Bayesian filtering.[1][2]
It also exposes tuning as a separate layer. Choosing \(\lambda\) trades historical information against responsiveness to change; choosing square-root or QR realization trades computational and stability properties. These decisions modify how the pattern operates without defining whether the pattern exists.[2]
Abstract Reasoning¶
For real regressors, a common objective at time \(t\) is \(J_t(w)=\sum_{i=1}^{t}\lambda^{t-i}(d_i-u_i^{T}w)^2\) plus any explicitly initialized regularizing term, with \(0<\lambda\leq1\). One conventional update uses a matrix \(P_{t-1}\) representing inverse weighted input information and computes \(k_t=P_{t-1}u_t/(\lambda+u_t^{T}P_{t-1}u_t)\). Then \(w_t=w_{t-1}+k_t(d_t-u_t^{T}w_{t-1})\), with a coupled rank-one-style update of \(P_t\). Different complex-signal conventions or factorized implementations write equivalent updates differently.[1][2]
This algebra explains why the gain is not a fixed learning rate: it depends on the new regressor relative to accumulated input geometry. It also explains why the old data need not be read again. But if \(P_{t-1}\) does not represent the intended historical objective—for example through a different initialization or numerical corruption—the update cannot simply be called the exact minimizer of an unqualified batch problem.[1][2]
Knowledge Transfer¶
In FIR identification, \(u_t\) is a sample window, \(d_t\) an unknown filter's output and \(w\) tap coefficients. In the valve example, \(u_t\) is a two-component regressor containing input and a constant, \(d_t\) measured valve response and \(w\) gain and offset. The interpretation changes while the least-squares state and prediction-error correction relation remains.[3][4]
Transfer does not extend to every online learner that updates a parameter vector. The method must maintain the particular quadratic-fit information needed for an RLS gain; otherwise “recursive least squares” is only a loose analogy.
Examples¶
FIR system identification. An unknown low-pass FIR filter maps input to desired output. The linear problem is to fit tap coefficients to that input-output stream. The maintained state is current taps plus inverse-correlation or factorized information. At each new sample, the predicted output differs from the observed target; the error-gain correction adjusts taps and matrix state. The official example's forgetting-factor control probes tracking but is not necessary to identify RLS.[3][1]
Mapped back: regressor, response, state and adaptive correction are all present; a graph of convergence is evidence about this realization, not the definition.
Valve gain and offset. Within the example's above-dead-band segment, \(y(t)=k(t)u(t)+b(t)+e(t)\) has regressor \([u(t),1]\) and coefficient pair \([k,b]\). Each new input/output pair in that segment supplies the linear residual. An RLS estimator retains the matrix information and applies a gain-weighted correction to the pair. The official example uses this to estimate changing parameters online, without treating the linear law as valid in the valve's dead band.[4][2]
Mapped back: the same update relation survives when filter taps become physical-model parameters; one should not infer an FIR waveform from that transfer.
Negative boundary: fixed-step LMS. An LMS filter may also adjust linear coefficients from each prediction error, but its gradient step does not maintain the RLS inverse-correlation or equivalent quadratic-fit state. Sharing a residual and sequential timing does not establish identity.[2]
Structural Tensions¶
- Tracking drift versus preserving history. \(\lambda<1\) discounts old evidence and may track changing coefficients, while \(\lambda=1\) preserves equal sample weights. Aggressive forgetting can make a fit noisy or fail to converge in the official FIR simulation. Diagnostic: Is the system changing, and does the chosen \(\lambda\) improve held-out tracking without destabilizing the estimate?[2][3]
- Adaptive gain versus update cost. Matrix-shaped gain can improve convergence under suitable conditions compared with simple gradient updates, but maintaining it costs more computation and numerical care. Diagnostic: What are the parameter dimension, input excitation, arithmetic precision and observed error trajectories?[2]
- Algebraic equivalence versus numerical reliability. Direct inverse-state equations express the objective cleanly, while factorized updates may preserve positive definiteness better under finite precision. Diagnostic: Does a batch check agree within tolerance, and does the maintained information state remain valid?[1][2]
Structural–Framed Character¶
Recursive Least-Squares Filter is mixed-structural, leaning formal: the sequential quadratic-fit update can be derived, while the linear model, weights and initialization determine what is estimated. Its evaluative weight is limited; RLS names an update method, not a promise of rapid convergence or a physically correct model. It is human-practice-bound as a designed algorithm deployed on chosen observations, though its objective relation is mathematically testable after those choices. Its institutional origin is adaptive filtering and system-identification research, not a standard that can confer correctness by labeling software “RLS.” Its vocabulary travel reaches acoustic, control and identification problems when the same sequential least-squares adaptation is performed. Import versus recognition requires a maintained fit state and recursive least-squares update, not merely any predictor that changes over time.
Live Estimation is a conceptual broad neighbor, but its present decision-purpose and uncertainty language was not judged a strict genus for all RLS filtering uses. A possible future-prime candidate is recursive maintenance of an optimization solution under arriving evidence; it is not an asserted graph node. Its character: a formally specified adaptive linear-estimation algorithm whose role pattern crosses applications but whose identity remains in sequential least-squares computation.
Structural Core vs. Domain Accent¶
The prime-bar question is which part of an adaptive update is truly portable.
What is skeletal. A solution to an objective can be updated as new observations arrive rather than refit from scratch. This recursive-maintenance relation is an explicit future-prime candidate, not a live strict parent. Linear Least Squares identifies the optimization problem that RLS addresses, while Estimation supplies a related broader purpose; problem, algorithm and purpose must not be collapsed into one is-a edge.
What is domain-bound. The filter maintains a state for a weighted quadratic fit of a linear observation model and uses a gain-like correction from the new residual. Remove the least-squares objective or recursive state update and an arbitrary adaptive predictor is not RLS. Equal weighting, exponential forgetting and sliding windows require distinct stated rules; a chosen \(\lambda\) is not universally less than one. FIR taps, acoustic channels and a valve's active-range line fit are application accents, while square-root versus direct covariance representation is an implementation choice with numerical consequences.
Why this is not a prime. Recursive optimization may have wider reach, but that skeleton requires separate admission. The named RLS filter is recognized where the sequential linear least-squares relation is actually implemented. An organization that revises estimates after reports arrive only resembles RLS unless it maintains the objective and update law. Thus the technical identity remains domain-specific and provisionally unparented, not a general synonym for learning from feedback.
Instantiates / Related Primes¶
The live Linear Least Squares (Linear least squares) names a mathematical optimization problem that RLS recursively addresses, but problem and algorithm are not automatically a strict is-a relation. Recursive Bayesian Estimation (Recursive Bayesian estimation) carries a posterior; its similar-looking gain in special cases is not the generic RLS objective. The live prime Estimation is conceptually broader but currently includes decision-purpose and uncertainty language not required by every RLS filtering instance. No weak topical edge is promoted to a strict DAG assertion.
Neighborhood in Abstraction Space¶
Recursive Least Squares Filter sits in a sparse region of the domain-specific corpus (78th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.
Family — Statistical Learning & Model Failure Modes (41 abstractions)
Nearest neighbors
- Structural Risk Minimization — 0.83
- Particle Filter — 0.83
- Lifting Scheme — 0.83
- Leabra — 0.82
- FWL theorem — 0.82
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
A batch least-squares solver can solve the same objective without a recursive state update. LMS can process the same samples with a fixed or normalized gradient step but not the same inverse-information recursion. Kalman filtering can coincide algebraically under particular linear-Gaussian and parameter-dynamics assumptions; those assumptions are not silently part of all RLS filters. Exponential forgetting is a configurable weighting choice, not proof that the data-generating system truly varies.[2]
References¶
[1] MathWorks, “dsp.RLSFilter—Compute output, error and coefficients using recursive least squares (RLS) algorithm”, Algorithms, forgetting-factor and initialization documentation, directly checked. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k
[2] MathWorks, “Recursive Algorithms for Online Parameter Estimation”, Forgetting Factor, Kalman Filter and gradient sections, directly checked. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u
[3] MathWorks, “System Identification Using RLS Adaptive Filtering”, Introduction and MATLAB Simulation, directly checked. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g
[4] MathWorks, “Line Fitting with Online Recursive Least Squares Estimation”, valve model and online parameter estimation example, directly checked. registry ↩a ↩b ↩c