Gamma filter
Model Families AIMech:GammaFilter DRAFT
Aliases: Adaptive gamma filter, ADALINE(μ)
A linear adaptive filter family that combines a direct input tap with weighted cascaded first-order gamma-memory traces, using one shared memory coefficient to control the restricted feedback.
Identity Notes & Concept Scope
Compared all 442 pinned AIO and 584 native IDs, names, aliases and scoped synonyms; no normalized identity match. Ignored- and hidden-inclusive fixed-string searches of the recovery, older discovery and original Documents checkouts returned 5, 0, 0 lines. All distinct match contents were inspected. Semantic matching covered 110 AIO and 251 native definitions, with unchanged preceding rows verified and new matches read. 4 full annotated AIO node/edge/generated contexts and 4 full native records were inspected. Matches point to the pending Gamma-filter lead and Gamma neural-model context. This source-defined linear adaptive filter has a direct input tap and a weighted sum of cascaded gamma-memory traces; the broader Gamma neural architecture allows nonlinear spatial maps and neural couplings. The independently named ADALINE(mu) filter is a narrower model family, not a renamed generic ADALINE or the entire Gamma neural family. LinearStateSpaceLayer concerns a layer based on discretized continuous state-space systems; PseudoLinearRegressionEstimation is an estimation method and IteratedWienerFilter is a particular iterative denoising procedure. None supplies this exact identity. Fixed-memory-coefficient fitting is linear in tap weights, while joint memory-coefficient adaptation is not linear in all parameters, motivating only a related LinearRegression mapping.
Claims & Grounded Evidence
6 Claims
Mechanism
#linear-filter-recurrence
For order parameter K, input u[n] and shared coefficient mu, define x0[n]=u[n] and update xk[n]=(1-mu)*xk[n-1]+mu*x(k-1)[n-1] simultaneously for k=1,...,K. Output y[n]=sum_(k=0..K) wk*xk[n] includes the direct w0 input tap. The transfer between adjacent stages is G(z)=mu/(z-(1-mu)); the output transfer is sum wk*G(z)^k. This parameterization has K+1 tap weights and one shared memory coefficient, hence K+2 adjustable scalars. It is a linear input-output filter for fixed coefficients, although its parameterization is not jointly linear in weights and mu.
Limitation
#fixed-memory-stability
For fixed real mu the memory-stage pole is 1-mu, strictly inside the unit circle when 0<mu<2. At mu=1 the stages form a tapped delay line and the filter becomes FIR. At mu=0 no new input enters the memory stages and their initial values remain frozen; mu=2 permits resonance from a bounded alternating input. Degenerate weights or cancellations can lower the output transfer order, so this stage condition is not an unconditional characterization of every degenerate output transfer. It also does not guarantee convergence or stability during unconstrained coefficient adaptation.
Mechanism
#dispersive-memory
With zero initial state and 0<mu<1, the kth-stage impulse response is binomial(n-1,k-1)*mu^k*(1-mu)^(n-k) for n>=k. It has unit sum and mean lag k/mu, so K/mu describes a mean depth for the last stage, not a hard memory horizon or an exact delayed sample. For 1<mu<2 the response alternates in sign and is not a probability distribution. The gamma-coordinate transformation makes the transfer polynomial in an alternative delay variable; that does not make the ordinary-time impulse response finite for general mu.
Mechanism
#block-and-online-adaptation
For a fixed-parameter block with half-squared error and e[n]=d[n]-y[n], gradient descent uses weight increments proportional to sum_n e[n]*xk[n] and a memory-coefficient increment proportional to sum_n e[n]*sum_k wk*alpha_k[n]. For parameter-independent zero initial state, alpha0=0 and alpha_k[n]=(1-mu)*alpha_k[n-1]+mu*alpha_(k-1)[n-1]+x_(k-1)[n-1]-xk[n-1]. Other initial-state conventions require corresponding initial sensitivities. The paper labels its samplewise updates a local-in-time approximation; fixed-parameter derivative identities do not establish exact gradients under arbitrary parameter drift. The direct recurrence uses O(K) work per step, without a universal training-time or convergence guarantee.
Limitation
#normal-equations-and-nonconvexity
For fixed mu in a stationary finite-second-moment setting, the tap-weight MSE normal equation is Rw=p with R=E[xx^T] and p=E[d*x]; uniqueness additionally needs suitable rank, such as positive-definite R. Direct differentiation gives the memory-coefficient stationarity condition E[e*sum_k wk*alpha_k]=0. A total derivative of R includes both E[alpha*x^T] and E[x*alpha^T], so a one-sided cross-moment must not be used as that full derivative. Joint adaptation in mu and the tap weights is not generally convex; the suggested mu=1 initialization is a heuristic, not a global optimum guarantee.
Scope
#system-identification-evidence
The reported experiments concern identification of two third-order elliptic low-pass systems. They compare analytic fixed-mu normal-equation solutions and samplewise tap-weight LMS results across selected mu values and orders. Such comparisons do not demonstrate automatic convergence of learned mu on arbitrary problems or universal superiority over FIR filters. Additional signal applications mentioned at the end of SectionIV are deferred to later publication and are not detailed experiments here.
⚠️ Curated Limitations & Negative Results
- This record covers the linear adaptive Gamma filter and its shared-coefficient parameterization, not every nonlinear Gamma neural network or arbitrary restricted-feedback filter.
- The fixed-stage stability interval is distinct from adaptation stability. Initial states, coefficient constraints, loss, learning rate and block or online update convention affect the actual procedure.
- The source notation beside Equation23 equates a total correlation-matrix derivative with a one-sided cross-moment. The record uses direct differentiation instead; it does not infer a defect in uninspected author code.
- The experiments evaluate selected filters and coefficient settings. No universal statistical efficiency, successful memory-order selection, adaptive global convergence or hardware speed claim follows.
- All eight primary PDF pages were read. The printed issue gives February1993 although inspected institutional metadata lists1992. Constructed recurrence and derivative checks involve no fitting; no author code, model training or empirical replication was performed.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:LinearRegression
Linear Regression
|
related_match | At fixed mu, fitting the linear combination of gamma-memory traces is a linear least-squares problem. The record identifies an adaptive dynamical filter family, while the pinned AIO class identifies regression analysis. Joint mu adaptation is nonlinear in the parameters, so neither exact identity nor a broad is-a mapping is asserted. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
principe1993
|
Jose C. Principe, Bert de Vries, Pedro G. de Oliveira
|
primary paper
full_text
|
ID: https://doi.org/10.1109/78.193206
sha256:0c4cf590b91e…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-13T07:11:25Z
Updated: 2026-10-13T07:13:52Z
Source File:
data/curated/GammaFilter.yaml