Set-Membership Affine Projection Algorithm

• Set-Membership Affine Projection Algorithm is an adaptive filtering method that updates filter coefficients only when the error exceeds a preset bound, reusing multiple recent data vectors for enhanced convergence.
• It applies a minimum-disturbance update strategy using projection geometry to satisfy posterior error constraints while maintaining robustness and preventing divergence.
• Design parameters like projection order, adaptive error bounds, and regularization are tuned to balance fast convergence, computational cost, and performance in interference rejection.

Set-Membership Affine Projection Algorithm (SM-AP) is a set-membership adaptive filtering method that combines affine projection with bounded-error feasibility constraints. Rather than updating the coefficient vector at every iteration, it updates only when the current estimation error violates a prescribed bound, and the active update is obtained as the minimum-disturbance correction that satisfies posterior error constraints over a block of recent data. In the literature, SM-AP is positioned between single-vector set-membership NLMS-type methods and RLS-type set-membership schemes: it reuses multiple recent regressors to accelerate convergence while retaining the data-selective update mechanism characteristic of set-membership filtering (Shabaani, 2020, Lamare et al., 2013).

1. Set-membership formulation and the role of affine projection

The set-membership viewpoint is organized around feasibility sets induced by an error bound. In the standard framework, the feasibility set is written as

$$\Theta=\bigcap_{(x,d)\in S}\{w\in\mathbb{C}^N:\ |d-x^T w|\le \alpha\},$$

the instantaneous constraint set is

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$

and the membership set is

$\Psi_n=\bigcap_{i=1}^n H_i.$

The key operational consequence is that if the current error already satisfies the bound, there is no need to update. This gives set-membership algorithms their data-selective character (Guo et al., 18 May 2026).

Within that framework, affine projection introduces data reuse. Instead of relying only on the current regressor, the algorithm uses a block of recent data vectors. In the DS-CDMA treatment of set-membership adaptive algorithms, SM-AP is described as a generalized case of the SM-NLMS where $P$ data vectors are used to increase the convergence speed. This places SM-AP between one-vector stochastic-gradient updates and least-squares or RLS-type recursions: it preserves the bounded-error philosophy of set-membership filtering while exploiting the fast convergence associated with projection over multiple recent regressors (Lamare et al., 2013).

A central conceptual distinction follows from this formulation. Classical point-estimation adaptive filters seek a single point solution, whereas set-membership methods define a feasibility set of acceptable solutions and update only when the current estimate violates an error bound. For SM-AP, the affine projection machinery specifies how the new coefficient vector is selected inside, or on the boundary of, that feasibility region (Shabaani, 2020).

2. Canonical SM-AP recursion

In the linear real-valued formulation used for robustness analysis, the desired signal is modeled as

$d(k)=w_o^T x(k)+n(k),$

where $w_o$ is the unknown optimal system, $x(k)$ is the input vector, and $n(k)$ is the noise or uncertainty term. The adaptive filter output is $y(k)=x^T(k)w(k)$, the scalar estimation error is $e(k)=d(k)-y(k)$, and the coefficient error vector is

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$0

For projection order $$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$1, the data matrix is

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$2

with associated error, desired-signal, noise, and constraint vectors collected over the same block (Shabaani, 2020).

The exact recursive SM-AP update is

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$3

where

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$4

assuming $$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$5 is invertible. The same recursion can be written with the indicator function

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$6

as

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$7

The update decision is therefore purely data selective: update if $$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$8, and otherwise leave the coefficient vector unchanged (Shabaani, 2020).

In the complex DS-CDMA formulation, the same structure appears with a block observation matrix

$$H_n=\{w\in\mathbb{C}^N:\ |d(n)-x^T(n)w|\le \alpha\},$$9

and the practical SM-AP recursion becomes

$\Psi_n=\bigcap_{i=1}^n H_i.$0

with

$\Psi_n=\bigcap_{i=1}^n H_i.$1

Here $\Psi_n=\bigcap_{i=1}^n H_i.$2, $\Psi_n=\bigcap_{i=1}^n H_i.$3 is the projection order, and $\Psi_n=\bigcap_{i=1}^n H_i.$4 is a small regularization constant (Lamare et al., 2013).

3. Minimum-disturbance derivation and projection geometry

The projection interpretation of SM-AP is explicit. Whenever an update occurs, $\Psi_n=\bigcap_{i=1}^n H_i.$5 is the solution of the constrained optimization problem

$\Psi_n=\bigcap_{i=1}^n H_i.$6

Thus, among all coefficient vectors satisfying the posterior error constraint, SM-AP picks the one with minimum Euclidean change from the current estimate. This minimum-disturbance principle is the defining geometric mechanism of the algorithm (Shabaani, 2020).

The DS-CDMA derivation makes the same structure explicit in block form: $\Psi_n=\bigcap_{i=1}^n H_i.$7 Using Lagrange multipliers yields

$\Psi_n=\bigcap_{i=1}^n H_i.$8

To obtain the set-membership affine projection recursion, the boundary is chosen so that

$\Psi_n=\bigcap_{i=1}^n H_i.$9

while the $P$0 older a posteriori errors are kept constant. This is what turns the general constrained solution into the practical data-selective SM-AP update (Lamare et al., 2013).

A common source of confusion is the word “projection.” The projection-operator literature distinguishes between projection of a point onto a set and modification of an update direction so that a parameter trajectory remains feasible. That distinction is directly relevant to adaptive filtering: the projection operator in adaptive systems is a direction-modification rule, not the classical Euclidean projector $P$1. This suggests that SM-AP should be understood as a feasibility-preserving adaptive law built from projection geometry, rather than only as a matrix formula (Lavretsky et al., 2011).

4. Stability, robustness, and non-divergence

The principal stability result currently available for the linear SM-AP algorithm is an $P$2-stability analysis. Starting from the coefficient-error recursion, the paper derives the energy identity

$P$3

and, at update times,

$P$4

This leads to the paper’s local robustness criterion: the one-step energy contracts if

$P$5

If no update occurs, then

$P$6

These statements are the basis of the local robustness theorem (Shabaani, 2020).

Global robustness is obtained by summing the local inequalities over the update times. If the constraint vector is chosen so that

$P$7

for every updating iteration, then

$P$8

The paper also gives an existence result: if

$P$9

then the robustness condition always holds. The same paper immediately notes that this choice is not practical because $d(k)=w_o^T x(k)+n(k),$0 is not observable (Shabaani, 2020).

The strongest claim of that analysis is distinct from global robustness. The paper argues that SM-AP never diverges, no matter how its parameters are chosen, because each active update solves a bounded posterior-error projection problem. In that reading, poor choices of the constraint vector may weaken robustness and worsen performance, but do not produce divergence. This distinction between conditional $d(k)=w_o^T x(k)+n(k),$1-robustness and parameter-independent non-divergence is central to the theoretical interpretation of SM-AP (Shabaani, 2020).

5. Design parameters, adaptive bounds, and application-specific behavior

The main design parameters of SM-AP are the projection order, the error bound, and the regularization. The projection order $d(k)=w_o^T x(k)+n(k),$2 determines how many recent data vectors are reused; larger $d(k)=w_o^T x(k)+n(k),$3 increases convergence speed but also increases computational cost because the recursion contains a $d(k)=w_o^T x(k)+n(k),$4 matrix inverse. The regularization $d(k)=w_o^T x(k)+n(k),$5 in

$d(k)=w_o^T x(k)+n(k),$6

is included explicitly for improving robustness (Lamare et al., 2013).

A major practical issue is specification of the error bound. In the DS-CDMA work, the fixed bound $d(k)=w_o^T x(k)+n(k),$7 is replaced by time-varying bounds. The Parameter Dependent Bound (PDB) is

$d(k)=w_o^T x(k)+n(k),$8

and the Parameter and Interference Dependent Bound (PIDB) is

$d(k)=w_o^T x(k)+n(k),$9

Because the SM-AP step size is

$w_o$0

a time-varying bound modifies both the update decision and the update magnitude. The same paper states that the SM-AP complexity is

$w_o$1

so the update rate $w_o$2 directly scales the effective cost (Lamare et al., 2013).

The reported DS-CDMA results attribute substantial gains to these adaptive bounds. For AP algorithms with $w_o$3, the PIDB mechanism gives the best BER performance across scenarios; the proposed PIDB mechanism with SM-AP and $w_o$4 can save up to $w_o$5 compared to PDB and up to $w_o$6 compared to SM-AP with fixed bounds for the same BER performance, and the PIDB approach can accommodate up to 4 more users than the PDB technique for the same BER. The interference-power estimator that supports PIDB tracks actual interference power very accurately, with average estimation error not exceeding about $w_o$7 of the estimated power level (Lamare et al., 2013).

The 2020 stability paper complements those results with explicit update-rate measurements under different constraint-vector choices. In system-identification simulations, the mean update rates for SM-AP were $w_o$8 for the general constraint vector, $w_o$9 for SC-CV, and $x(k)$0 for the noise CV. The same simulations reported that SM-AP with the practical SC-CV choice sharply reduced the number of iterations for which $x(k)$1, and that the idealized $x(k)$2 choice made $x(k)$3 monotone decreasing (Shabaani, 2020).

6. Widely-linear and kernelized generalizations

A major recent generalization is the Augmented Set-Membership Affine Projection Algorithm (ASM-APA), developed for complex, second-order noncircular, and highly colored input signals. In that setting, the strictly linear model is insufficient, and the widely-linear filter output is

$x(k)$4

The ASM-APA derivation combines the Augmented Affine Projection Algorithm (AAPA) with set-membership filtering, preserving the augmented affine projection geometry while updating only when the error exceeds a prescribed bound. The final recursion is

$x(k)$5

$x(k)$6

with

$x(k)$7

The paper emphasizes that ASM-APA reduces the average inversion burden by performing updates only on informative samples; it does not eliminate per-update inversion algebraically. In the reported experiments, ASM-APA consistently achieves smaller steady-state error than AAPA, its update rate remains around $x(k)$8–$x(k)$9 in the highly correlated system-identification setting, and in stereophonic acoustic echo cancellation it attains faster convergence, lower steady-state inaccuracy, and an update rate of only $n(k)$0 compared with AAPA (Guo et al., 18 May 2026).

A nonlinear generalization appears in the set-membership kernelized affine projection algorithm, denoted SM-KAP or SM-KAPA. In reproducing kernel Hilbert space form, the algorithm solves

$n(k)$1

which yields

$n(k)$2

with

$n(k)$3

This is the RKHS counterpart of linear SM-AP: the linear data matrix is replaced by the feature-space matrix, and the correlation matrix is replaced by the kernel Gram matrix. In the reported time-series prediction experiments, SM-KAPA achieved test MSE $n(k)$4 on Mackey-Glass and $n(k)$5 on laser data, with $n(k)$6, Gaussian kernel bandwidth $n(k)$7, and error bound $n(k)$8 (Lamare et al., 2017).

These variants preserve the same structural core: bounded-error feasibility, minimum-disturbance projection, and data-selective updating. What changes is the geometry of the regressor space—real linear, widely-linear augmented complex, or RKHS—together with the form of the projection matrix and the application domain (Guo et al., 18 May 2026, Lamare et al., 2017).