Affine Filtering Measurements

• Affine filtering measurements are techniques that incorporate predetermined bias terms to convert symmetric, ambiguous measurements into ones with clear, interpretable structure.
• They enable exact phase retrieval and improve identifiability by transforming classical models into affine formulations, reducing sample complexity in both real and complex settings.
• The approach finds applications in sparse recovery, low-rank matrix reconstruction, Bayesian filtering, image processing, communications, and quantum decoding, each exploiting the affine structure for enhanced robustness.

Affine filtering measurements denote a family of measurement and filtering constructions in which linear filter outputs are modified by known affine terms, or in which inference is organized around affine maps, affine subspaces, or local affine models. In the phase-retrieval literature, the canonical measurement is

$$y_j = |\langle a_j,x\rangle + b_j|, \qquad j=1,\ldots,m,$$

and the decisive structural feature is that the offsets $b_j$ remove the global unimodular ambiguity of standard phase retrieval, so that $x$ can be recovered exactly rather than only up to a global phase (Gao et al., 2016). The same expression also appears in sparse phase retrieval, generalized affine quadratic measurements, low-rank matrix recovery, guided image filtering, Bayesian state estimation, quantum decoding, and affine filter-bank modulation, where the common thread is the use of affine structure to improve identifiability, conditioning, or interpretability (Yang et al., 2021, Huang et al., 2018, Wagner et al., 2015, Zhao et al., 5 Mar 2025, Wen et al., 2021, Mandal et al., 5 Jun 2026, Senger et al., 6 May 2025).

1. Canonical model: affine phase retrieval from biased linear filters

The basic affine phase-retrieval model fixes $a_j\in\mathbb{H}^d$ and $b_j\in\mathbb{H}$, with $\mathbb{H}=\mathbb{R}$ or $\mathbb{C}$, and observes either

$$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$

or

$$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$

Affine phase retrievability means that $M_{A,b}$ is injective on $b_j$0, and the paper shows that injectivity of $b_j$1 and $b_j$2 is equivalent (Gao et al., 2016).

This model can be rewritten as ordinary phase retrieval in one higher dimension. With

$b_j$3

one has

$b_j$4

The difference from classical phase retrieval is that the last coordinate of $b_j$5 is known to be $b_j$6, so the usual unimodular ambiguity is removed. This reduction clarifies why affine offsets act as known references and why exact signal recovery becomes possible.

A standard interpretation is to view each measurement as the magnitude of the output of a linear filter $b_j$7 applied to $b_j$8, followed by addition of a known bias $b_j$9 before magnitude extraction. In that sense, affine filtering measurements are magnitude-only measurements of biased filter outputs. The same interpretation extends to practical filter-bank design: in the real case, two distinct biases per coordinate suffice in the minimal construction, whereas in the complex case three noncollinear biases per coordinate suffice.

2. Injectivity criteria and measurement complexity

In the real case, the main equivalences are combinatorial, Jacobian, and bilinear. For $x$0 and $x$1, affine phase retrievability is equivalent to each of the following: for any $x$2 with $x$3, there exists $x$4 such that

$x$5

for any subset $x$6, if $x$7, then $x$8; and the Jacobian of $x$9 has rank $a_j\in\mathbb{H}^d$0 for all $a_j\in\mathbb{H}^d$1. In the complex case, the corresponding criterion is

$a_j\in\mathbb{H}^d$2

for some $a_j\in\mathbb{H}^d$3, and the real Jacobian must have rank $a_j\in\mathbb{H}^d$4 everywhere (Gao et al., 2016).

The minimal measurement counts differ sharply from standard phase retrieval. For real signals, $a_j\in\mathbb{H}^d$5 is necessary, and a generic $a_j\in\mathbb{H}^d$6 is affine phase retrievable when $a_j\in\mathbb{H}^d$7. For complex signals, $a_j\in\mathbb{H}^d$8 is necessary, and this bound is sharp: if $a_j\in\mathbb{H}^d$9 is nonsingular, $b_j\in\mathbb{H}$0, and each triple $b_j\in\mathbb{H}$1 is noncollinear in the complex plane, then $b_j\in\mathbb{H}$2 is affine phase retrievable with $b_j\in\mathbb{H}$3. Generic sufficiency in the complex case is proved for $b_j\in\mathbb{H}$4 (Gao et al., 2016).

A generalized version replaces scalar affine filters by vector-valued affine filters,

$b_j\in\mathbb{H}$5

with $b_j\in\mathbb{H}$6. This generalized affine phase retrieval has necessary bounds

$b_j\in\mathbb{H}$7

and these are tight when $b_j\in\mathbb{H}$8. The construction is blockwise: partition coordinates into $b_j\in\mathbb{H}$9 blocks, use $\mathbb{H}=\mathbb{R}$0 that extract one block, and choose $\mathbb{H}=\mathbb{R}$1 offsets in the real case or $\mathbb{H}=\mathbb{R}$2 offsets in the complex case so that the block is uniquely determined from its affine-norm samples (Huang et al., 2018).

3. Sparse affine sampling and ambiguity-free sparse recovery

Sparse affine sampling specializes the affine model to $\mathbb{H}=\mathbb{R}$3-sparse signals and magnitude-squared measurements

$\mathbb{H}=\mathbb{R}$4

The sensing matrix is sparse: each column $\mathbb{H}=\mathbb{R}$5 has support $\mathbb{H}=\mathbb{R}$6 of fixed size $\mathbb{H}=\mathbb{R}$7, with limited overlap $\mathbb{H}=\mathbb{R}$8. The method is explicitly two-stage: support identification, followed by exact or stable recovery of the nonzero entries (Yang et al., 2021).

In the noise-free case, support recovery uses a counting rule relative to the bias-only level $\mathbb{H}=\mathbb{R}$9. Under the non-collinearity condition that for any fixed $\mathbb{C}$0 and any distinct $\mathbb{C}$1, the three centers $\mathbb{C}$2 are not collinear, and under

$\mathbb{C}$3

one can choose

$\mathbb{C}$4

and recover the support exactly. Once the support is known, each active scalar $\mathbb{C}$5 is recovered from the intersection of three circles

$\mathbb{C}$6

or equivalently by eliminating $\mathbb{C}$7 from pairwise differences.

The framework extends to sparse outliers and bounded dense noise. For sparse outliers of support size $\mathbb{C}$8, exact support identification remains possible if

$\mathbb{C}$9

For bounded noise, support identification uses the thresholded deviation $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$0, and subsequent recovery is by linear least squares on pairwise-difference equations. The resulting error bound depends on the dispersion and non-collinearity of the centered reference points, and the paper shows that near-optimal performance can be achieved with high probability by drawing the nonzero entries of the sparse sensing matrix and the bias vector uniformly on a complex circle (Yang et al., 2021).

This sparse theory preserves the central geometric idea of affine filtering measurements: the bias vector creates reference-centered circles, so phase ambiguity is replaced by an explicitly solvable geometry.

4. Stability, topology, and reconstruction algorithms

Affine phase retrievability has compact-set stability but not global bi-Lipschitzness. If $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$1 is affine phase retrievable and $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$2 is compact, then there exist positive constants $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$3 such that for all $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$4,

$$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$5

and

$$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$6

Neither $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$7 nor $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$8 is bi-Lipschitz on all of $$M_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|,\ldots,|\langle a_m,x\rangle+b_m|\big)$$9 (Gao et al., 2016).

A notable topological difference from standard phase retrieval is non-openness. In the classical case, phase-retrievable measurement sets are open. In the affine case, the set of affine phase retrievable $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$0 is not open in $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$1: there are injective designs for which arbitrarily small perturbations of $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$2 destroy injectivity. This non-openness appears in both the real and complex settings and has direct implications for fabrication, calibration, and robustness margins.

Algorithmically, the augmentation

$$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$3

connects affine phase retrieval to classical solvers such as PhaseLift, matrix completion, Wirtinger flow, and alternating minimization, now with the anchor constraint that the last coordinate of $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$4 equals $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$5 (Gao et al., 2016). In the generalized norm model, the lifting becomes

$$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$6

with $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$7 determined by $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$8 and $$M^2_{A,b}(x)=\big(|\langle a_1,x\rangle+b_1|^2,\ldots,|\langle a_m,x\rangle+b_m|^2\big).$$9. This suggests SDP-based and nonconvex least-squares reconstructions, although the cited work is focused on identifiability rather than full algorithmic analysis (Huang et al., 2018).

A frequent misconception is that adding affine offsets automatically improves robustness. The theory supports a narrower statement: offsets remove global-phase ambiguity and can reduce measurement counts, especially in the complex setting, but injective affine designs do not form an open set, so robustness to perturbation is a separate design issue.

5. Affine measurement models beyond phase retrieval

Low-rank recovery provides a matrix-valued affine counterpart. In row-and-column affine measurements one observes

$M_{A,b}$0

where $M_{A,b}$1 is rank $M_{A,b}$2. The proposed SVLS algorithm combines an SVD on one measurement block with least squares on the other; in the noiseless Gaussian row/column design, exact recovery holds with probability $M_{A,b}$3 when $M_{A,b}$4, and in the noisy case the Frobenius error satisfies a bound of the form

$M_{A,b}$5

with high probability (Wagner et al., 2015). A related affine-rank-minimization formulation is

$M_{A,b}$6

for which the BARM algorithm optimizes

$M_{A,b}$7

and is empirically reported to succeed at the degrees-of-freedom limit $M_{A,b}$8 across matrix completion, general affine operators, and image rectification (Xin et al., 2014).

In Bayesian filtering, the phrase appears in a different but structurally related sense. Standard Gaussian filters represent the posterior belief by a Gaussian whose mean is an affine function of the measurement, and the Feature Gaussian Filter relaxes this by making the mean affine in a nonlinear feature map $M_{A,b}$9:

$b_j$00

The affine-mapping based variational Ensemble Kalman filter likewise searches for an affine map $b_j$01 that pushes the prior ensemble toward the posterior by minimizing a KL divergence, with exact agreement with the Kalman update in the linear-Gaussian case (Wüthrich et al., 2015, Wen et al., 2021).

Piecewise-affine and affine-process state estimation use yet another meaning. For piecewise affine state-space models, each filtering recursion yields a mixture of truncated normal distributions that is approximated by a single normal via moment matching (Rui et al., 2016). For affine processes observed in additive white noise, the linearized filtering functional replaces the intractable quadratic observation term by an affine linearization around a macro location, and the conditional characteristic function is then computed through generalized Riccati equations driven by the observation path (Gonon et al., 2018).

In image filtering, the original guided image filter is based on the local affine model

$b_j$02

with coefficients estimated from local means, variances, and covariances. The Gaussian highpass guided-filtering formulation replaces this two-parameter local affine model by

$b_j$03

so that guidance structure enters through a Gaussian highpass component and the input contributes a Gaussian-smoothed base (Zhao et al., 5 Mar 2025). The paper reports that these PM-GF variants outperform their LAM counterparts across edge-aware smoothing, denoising, detail enhancement, HDR tone mapping, dehazing, and texture removal, while also noting that inconsistent guidance structures can introduce substantial errors.

6. Communications and quantum decoding interpretations

In high-mobility communications, Affine Filter Bank Modulation combines a filter-bank waveform with DAFT precoding. The DAFT is

$b_j$04

and the resulting waveform uses filtered chirps, with only the first and last $b_j$05 subcarriers active to preserve complex orthogonality after filtering and compensation. Analytical and simulation results show quasi-orthogonality similar to AFDM in doubly-dispersive channels, PAPR levels $b_j$06 dB lower, and OOBE as low as $b_j$07 dB when PHYDYAS prototype filters are used (Senger et al., 6 May 2025). A later SIR analysis shows that under MMSE equalization the filtered time-domain detector exhibits a cancellation between channel-induced interference and orthogonality approximation error that does not occur in the affine domain, producing large SIR gains and explaining the associated BER improvement (Senger et al., 26 Nov 2025).

In quantum decoding, affine filtering measurements are a structured form of unambiguous state discrimination for codeword-indexed pure-state families. The POVM is

$b_j$08

where each conclusive outcome identifies an affine coset

$b_j$09

and must satisfy the unambiguity constraint

$b_j$10

The optimal AFM design is an SDP that reduces to a linear program via character-based diagonalization, and the resulting decoder for local codes accumulates error-free linear equations from conclusive outcomes and then solves them by Gaussian elimination (Mandal et al., 5 Jun 2026).

For regular LDPC codes from Gallager ensembles on i.i.d. pure-state channels, the reported thresholds show consistent gains over symbol-wise USD and symbol-wise PGM-based decoding. For example, for $b_j$11 ensembles the AFM+GE threshold is approximately $b_j$12, compared with $b_j$13 for USD+GE and $b_j$14 for PGM+BP; similar improvements are listed for $b_j$15, $b_j$16, $b_j$17, and $b_j$18 ensembles (Mandal et al., 5 Jun 2026). In this setting, affine filtering measurements are not biased magnitude samples but code-aware, affine-subspace-valued measurement outcomes.

Across these disparate literatures, the phrase retains a stable structural meaning. An affine filtering measurement does not merely record a linear response; it augments that response by a bias, an affine map, an affine subspace label, or a local affine model. The practical consequences vary by domain—ambiguity-free recovery in phase retrieval, reduced sample complexity in generalized affine sensing, efficient low-rank reconstruction, more expressive Bayesian updates, interpretable structure transfer in image filtering, filtered-chirp robustness in doubly-dispersive channels, and code-aware quantum decoding—but in each case the affine structure is used to convert an otherwise symmetric or weakly informative measurement model into one with stronger identifiability or more useful downstream constraints.