M-FISHER: Multi-Domain Fisher Approaches

• M-FISHER is a context-dependent term describing multiple research frameworks that apply Fisher theory to industrial signals, sequential adaptation, optical measurements, spatial regression, and independence testing.
• The techniques employ diverse methodologies such as STFT sub-band modeling with teacher-student self-distillation, martingale-driven shift detection with Fisher-preconditioned updates, and optimization via Fisher-information matrices.
• Clear disambiguation is essential, as the term unifies distinct, non-interchangeable methods that leverage Fisher constructs in varying theoretical and applied contexts.

Searching arXiv for the provided topic and related papers. “M-FISHER” is not a single universally standardized term in the arXiv literature. It appears in at least four distinct senses: as an alias for the industrial foundation model FISHER for heterogeneous SCADA and industrial IoT signals (Fan et al., 22 Jul 2025); as a sequential test-time adaptation framework based on martingale-driven Fisher prompting (Khan et al., 4 Oct 2025); as shorthand for a multiparameter Fisher-information methodology for coherent diffraction measurements (Verreussel et al., 31 May 2026); and as an informal label for a Fisher-consistent redescending M-estimation framework in spatial scalar-on-function regression (Mutis et al., 1 May 2026). A further, phonetically related but technically separate usage is the multiscale Fisher’s independence test, also abbreviated in the source material as “M-FISHER” or MultiFIT (Gorsky et al., 2018). The term therefore denotes a family of unrelated research constructs linked primarily by the word “Fisher,” rather than a single theory, algorithm, or model family.

1. Terminological scope and disambiguation

The most immediate source of ambiguity is that the source literature uses “M-FISHER” in different domains with incompatible meanings. In industrial representation learning, the name refers to what the corresponding paper itself calls simply FISHER, namely “a Foundation model for multi-modal Industrial Signal compreHEnsive Representation” (Fan et al., 22 Jul 2025). The accompanying details explicitly state that “M-FISHER is our name for what in the paper is simply called FISHER,” identifying the “M” as a retrospective naming convention rather than the canonical title.

In sequential adaptation, “M-FISHER” names a framework for “sequential distribution shift detection and stable adaptation in streaming data,” combining an e-process detector with Fisher-preconditioned prompt updates (Khan et al., 4 Oct 2025). In optical metrology, “M-Fisher” denotes a “multiparameter Fisher-information” approach for coherent diffraction measurements, centered on scalar criteria defined on a Fisher-information matrix (Verreussel et al., 31 May 2026). In robust spatial statistics, “M-Fisher” refers to a “Fisher-consistent redescending M-estimation” construction for spatial scalar-on-function regression (Mutis et al., 1 May 2026). In nonparametric dependence testing, “M-FISHER” or MultiFIT denotes the “Multi-scale Fisher’s Independence Test for Multivariate Dependence” (Gorsky et al., 2018).

A plausible implication is that any technical discussion of “M-FISHER” should begin with domain disambiguation. Without that step, claims about architecture, asymptotic guarantees, or Fisher information can be incorrectly transferred across unrelated literatures.

2. M-FISHER as unified industrial signal foundation modeling

In the industrial-signal literature, M-FISHER is a unified foundation model motivated by the “M5 problem,” which comprises five axes of heterogeneity: “Multi-modal,” “Multi-sampling-rate,” “Multi-scale,” “Multi-task,” and “Mini-fault” (Fan et al., 22 Jul 2025). The signals explicitly listed are “sound, vibration, voltage, current, temperature, etc.,” and the motivating setting is the rapid deployment of SCADA systems together with the need to “effectively analyze industrial signals and detect abnormal states” (Fan et al., 22 Jul 2025).

The central claim is that these heterogeneous streams “can be modeled in a unified manner due to the intrinsic similarity” and that previous work treated them as siloed sub-problems with separate models, fixed sampling rates, and distinct preprocessing and deployment pipelines (Fan et al., 22 Jul 2025). The proposed representation begins by converting a raw 1D sensor stream $x[n]$ into an STFT spectrogram with fixed time-duration window $t_{\rm win}$ and hop $t_{\rm hop}$, so that for sample rate $f_s$ the FFT window length is $N=t_{\rm win}\cdot f_s$, the hop length is $H=t_{\rm hop}\cdot f_s$, and the frequency resolution is $\Delta f=f_s/N=1/t_{\rm win}$. The STFT is written as

$$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$

Rather than mel-filterbanks or log-mel, the model “retain[s] the raw linear-scale STFT,” because “mechanical faults often speak in high or harmonic frequencies that mel scales can obscure” (Fan et al., 22 Jul 2025). The frequency bins are then partitioned into contiguous sub-bands $B_1,\dots,B_K$ of fixed bandwidth $w=f_{\rm base}\cdot t_{\rm win}$, and the total representation is

$t_{\rm win}$0

When the sampling rate increases, “the extra Nyquist-band appears as extra $t_{\rm win}$1’s” (Fan et al., 22 Jul 2025). In practical batching, a common $t_{\rm win}$2 is randomly selected, each clip is resampled to that rate, the STFT is computed with shape $t_{\rm win}$3, the result is split into $t_{\rm win}$4 sub-bands of shape $t_{\rm win}$5, and the sub-band slices are stacked along the batch axis to obtain $t_{\rm win}$6 (Fan et al., 22 Jul 2025).

The pre-training architecture uses a teacher-student EMA framework with masked prediction self-supervision (Fan et al., 22 Jul 2025). The student consists of encoder $t_{\rm win}$7 and decoder $t_{\rm win}$8, while the teacher encoder $t_{\rm win}$9 shares the architecture of $t_{\rm hop}$0 and is updated by exponential moving average:

$t_{\rm hop}$1

with $t_{\rm hop}$2 close to 1. Each sub-band is tokenized into $t_{\rm hop}$3 patches, and “80% of patches are masked via ‘inverse block’ masking (as in EAT)” (Fan et al., 22 Jul 2025). The student produces a sub-band embedding $t_{\rm hop}$4 and patch reconstructions $t_{\rm hop}$5; the teacher produces $t_{\rm hop}$6 and $t_{\rm hop}$7 “by averaging its layer outputs and without gradient flow” (Fan et al., 22 Jul 2025). The losses are

$t_{\rm hop}$8

with total objective

$t_{\rm hop}$9

At inference, the model runs $f_s$0 on each sub-band, extracts the per-band [CLS] embedding, and concatenates the results into a final representation (Fan et al., 22 Jul 2025).

The downstream evaluation benchmark is RMIS, “Representation of M5 Industrial Signals,” an 18-dataset suite containing “Anomaly Detection (5 DCASE ‘unsupervised ASD’ tasks, 16 kHz sound, AUC metric)” and “Fault Diagnosis (13 sealed train–test splits across Sound, Vibration, Voltage, Current; macro-accuracy metric)” (Fan et al., 22 Jul 2025). The encoder is frozen; a memory bank of training-set embeddings is built; and inference uses k-NN with “$f_s$1 for ASD, $f_s$2 for FD” (Fan et al., 22 Jul 2025). Performance highlights reported in the source include: “FISHER-small (22 M parameters) obtains an overall RMIS score of 62.50%, exceeding all prior audio SSL baselines (AudioMAE, BEATs, EAT, CED, DaSheng) by at least +5.03 pp”; “On fault diagnosis alone, FISHER-tiny (5.5 M params) already beats every baseline by +9.24 pp”; “On anomaly detection, FISHER-small is runner-up only to BEATs by ≈0.8 pp, while simultaneously dominating the fault-diagnosis side”; and “Even FISHER-tiny outperforms baselines with 90 M+ parameters by +3.9 pp overall” (Fan et al., 22 Jul 2025).

The same work emphasizes scaling behavior. Three released sizes—“tiny (5.5 M), mini (10 M), small (22 M)”—are all “12-layer ViTs with patch size 16×16,” and RMIS performance “rises almost linearly” with size, whereas baselines “plateau or even dip beyond ~100 M params” (Fan et al., 22 Jul 2025). Pre-training uses “~17 k hours of general sound (AudioSet, FreeSound, Music4All),” but the paper notes that effective unique industrial signals are far fewer and that “further data diversity (new machines, modalities, conditions) is critical for scaling” (Fan et al., 22 Jul 2025). It also identifies “Test-Time Scaling” through higher-rate inputs: sub-band concatenation can exploit added frequency content “an inexpensive way to boost performance on high-rate data” (Fan et al., 22 Jul 2025).

3. M-FISHER as martingale-driven sequential test-time adaptation

A second use of the term defines M-FISHER as a method for “sequential distribution shift detection and stable adaptation in streaming data” (Khan et al., 4 Oct 2025). The framework begins with a non-conformity score

$f_s$3

where the first term measures how confident the predictive distribution is relative to a uniform guess and the second is a Mahalanobis distance between the current embedding and the training-set mean (Khan et al., 4 Oct 2025). Under no shift, these scores fluctuate around a calibration mean; under drift, they deviate systematically.

For sequential detection, the method constructs an exponential supermartingale from calibration statistics. Given held-out calibration scores $f_s$4, it computes

$f_s$5

and

$f_s$6

The e-process is then

$f_s$7

By Ville’s inequality,

$f_s$8

so setting $f_s$9 yields a time-uniform false-alarm bound of $N=t_{\rm win}\cdot f_s$0 (Khan et al., 4 Oct 2025). The stopping time is defined as

$N=t_{\rm win}\cdot f_s$1

Under a sustained shift after unknown time $N=t_{\rm win}\cdot f_s$2, the framework defines

$N=t_{\rm win}\cdot f_s$3

and the post-shift information-gain quantity

$N=t_{\rm win}\cdot f_s$4

The expected delay then satisfies the reported order bound

$N=t_{\rm win}\cdot f_s$5

and with $N=t_{\rm win}\cdot f_s$6 this becomes $N=t_{\rm win}\cdot f_s$7 (Khan et al., 4 Oct 2025). In the paper’s interpretation, a larger distributional shift in score space corresponds to larger $N=t_{\rm win}\cdot f_s$8 and therefore faster detection.

Adaptation is carried out by Fisher-preconditioned prompt updates. The prompt-level Fisher information is

$N=t_{\rm win}\cdot f_s$9

With instantaneous loss $H=t_{\rm hop}\cdot f_s$0, the Fisher-preconditioned gradient is

$H=t_{\rm hop}\cdot f_s$1

and the update rule is

$H=t_{\rm hop}\cdot f_s$2

The theoretical justification given is that small prompt perturbations satisfy the local approximation

$H=t_{\rm hop}\cdot f_s$3

so the update direction $H=t_{\rm hop}\cdot f_s$4 is the steepest descent direction in Fisher-Rao geometry, with the stated consequences of “parameterization-invariance” and “stability” (Khan et al., 4 Oct 2025).

Operationally, the full algorithm proceeds at each time step by computing $H=t_{\rm hop}\cdot f_s$5, updating $H=t_{\rm hop}\cdot f_s$6, and, if $H=t_{\rm hop}\cdot f_s$7, declaring shift and updating the prompt via the Fisher-preconditioned gradient (Khan et al., 4 Oct 2025). The source frames this as a principled separation between “when to update,” determined by the martingale trigger, and “how to update,” determined by the natural-gradient rule.

4. M-Fisher as multiparameter Fisher-information optimization in optical metrology

In coherent diffraction measurements, M-Fisher denotes a multiparameter Fisher-information framework for designing maximum-information input modes (Verreussel et al., 31 May 2026). The physical system is described by a scattering matrix

$H=t_{\rm hop}\cdot f_s$8

where $H=t_{\rm hop}\cdot f_s$9 is the parameter vector (Verreussel et al., 31 May 2026). Under heterodyne detection of all output modes, the Fisher-information matrix is

$\Delta f=f_s/N=1/t_{\rm win}$0

with Fisher-information operator

$\Delta f=f_s/N=1/t_{\rm win}$1

An equivalent coordinate form is

$\Delta f=f_s/N=1/t_{\rm win}$2

where $\Delta f=f_s/N=1/t_{\rm win}$3 (Verreussel et al., 31 May 2026).

Because entrywise optimization of $\Delta f=f_s/N=1/t_{\rm win}$4 is impractical, the paper introduces scalar optimality criteria: D-optimality, maximizing $\Delta f=f_s/N=1/t_{\rm win}$5; A-optimality, minimizing $\Delta f=f_s/N=1/t_{\rm win}$6; E-optimality, maximizing $\Delta f=f_s/N=1/t_{\rm win}$7; a scalar-CRLB objective, maximizing $\Delta f=f_s/N=1/t_{\rm win}$8 for a single parameter; and “Normalized trace,” defined as

$\Delta f=f_s/N=1/t_{\rm win}$9

These are further organized as special cases of a generalized mean of eigenvalues,

$$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$0

with D-, A-, and E-optimality corresponding, up to monotonic re-scaling, to $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$1 (Verreussel et al., 31 May 2026).

The maximum-information states depend on the criterion. For scalar-CRLB, one defines

$$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$2

diagonalizes it, and selects the top eigenvector $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$3 (Verreussel et al., 31 May 2026). For normalized trace, one forms

$$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$4

and again chooses the principal eigenvector (Verreussel et al., 31 May 2026). For D-, A-, and E-optimality, the paper states that “one must directly optimize” the chosen scalar criterion under the sphere constraint $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$5 using “standard gradient-descent or manifold-optimization routines” (Verreussel et al., 31 May 2026).

The framework also treats nuisance parameters. If $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$6 splits into parameters of interest and nuisance parameters, the Fisher matrix is partitioned into four blocks, and three strategies are considered: “Full-Fisher,” optimizing $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$7 on the full matrix; “Subblock,” optimizing $$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$8 as if nuisances were known; and “Partial-Fisher (Schur complement),” optimizing

$$X_{t,f}=\sum_{n=0}^{N-1}x[n]\cdot w[n-tH]\cdot e^{-j2\pi fn/N}.$$9

which yields the bound

$B_1,\dots,B_K$0

The paper reports that, via Pareto-front analysis, partial-Fisher for D- and A-criteria lies on the Pareto-optimal trade-off curve, while subblock can be slightly worse and full-Fisher is advantageous only when one also cares about the nuisance CRLBs (Verreussel et al., 31 May 2026).

The numerical example uses “two vertical columns of $B_1,\dots,B_K$1 point-dipole scatterers in a 2D scalar-wave simulation,” with the right column shifted horizontally by $B_1,\dots,B_K$2, vertically by $B_1,\dots,B_K$3, and rotated by $B_1,\dots,B_K$4 (Verreussel et al., 31 May 2026). The illumination basis contains “45 plane-wave illumination modes” within numerical NA $B_1,\dots,B_K$5, and partial derivatives are computed by symmetric difference with $B_1,\dots,B_K$6 units (Verreussel et al., 31 May 2026). The reported outcome is that “all multiparameter criteria (D, A, normalized-trace) substantially outperform plane waves or random fields, typically by an order of magnitude in CRLB,” and that D- and A-criteria yield joint CRLBs close to the single-parameter CRLBs (Verreussel et al., 31 May 2026).

5. M-Fisher as Fisher-consistent redescending M-estimation in spatial scalar-on-function regression

In spatial statistics, “M-Fisher” is used in the source material as shorthand for a Fisher-consistent redescending M-estimation framework for the spatial scalar-on-function regression model (Mutis et al., 1 May 2026). The model is defined over areal units $B_1,\dots,B_K$7, with scalar response $B_1,\dots,B_K$8 and functional predictor $B_1,\dots,B_K$9. With row-normalized spatial-weights matrix $w=f_{\rm base}\cdot t_{\rm win}$0, $w=f_{\rm base}\cdot t_{\rm win}$1, $w=f_{\rm base}\cdot t_{\rm win}$2, and spatial-autoregressive parameter $w=f_{\rm base}\cdot t_{\rm win}$3, the model is

$w=f_{\rm base}\cdot t_{\rm win}$4

where $w=f_{\rm base}\cdot t_{\rm win}$5 (Mutis et al., 1 May 2026). After robust functional PCA, the finite-dimensional form becomes

$w=f_{\rm base}\cdot t_{\rm win}$6

with robust principal-component expansion

$w=f_{\rm base}\cdot t_{\rm win}$7

and coefficient vector $w=f_{\rm base}\cdot t_{\rm win}$8 (Mutis et al., 1 May 2026).

The robust FPCA stage uses projection-pursuit RFPCA. For any unit-norm direction $w=f_{\rm base}\cdot t_{\rm win}$9, the univariate scores are $t_{\rm win}$00, with M-scale $t_{\rm win}$01 defined via a bounded even loss $t_{\rm win}$02 (Mutis et al., 1 May 2026). The principal directions are obtained by sequential maximization:

$t_{\rm win}$03

and for $t_{\rm win}$04,

$t_{\rm win}$05

Under an elliptical population law for $t_{\rm win}$06, these RFPCA functionals are stated to be Fisher-consistent for the true PCA basis up to a scale factor (Mutis et al., 1 May 2026).

The M-Fisher estimator is then defined from joint bias-corrected estimating equations. Let residuals be

$t_{\rm win}$07

With continuously differentiable even redescending loss $t_{\rm win}$08 and score $t_{\rm win}$09, the paper considers two specific losses: Danish and Andrews (Mutis et al., 1 May 2026). Because the spatial lag $t_{\rm win}$10 is endogenous, the naive M-equation for $t_{\rm win}$11 is biased, so the method introduces

$t_{\rm win}$12

where $t_{\rm win}$13 under the standardized $t_{\rm win}$14-law, with calibration constants “$t_{\rm win}$15 for Andrews, $t_{\rm win}$16 for Danish” (Mutis et al., 1 May 2026).

The estimator $t_{\rm win}$17 solves

$t_{\rm win}$18

$t_{\rm win}$19

$t_{\rm win}$20

The source interprets these as identifying $t_{\rm win}$21, $t_{\rm win}$22 in a “Fisher-consistent M-scale sense,” and $t_{\rm win}$23 with analytic correction (Mutis et al., 1 May 2026). Fisher consistency is expressed at the population level by the vanishing of the corresponding expectations at the true parameter $t_{\rm win}$24 (Mutis et al., 1 May 2026).

The asymptotic theory summarized in the source includes Fisher consistency, consistency via Z-estimation, asymptotic normality,

$t_{\rm win}$25

and the asymptotic distribution of the reconstructed slope function

$t_{\rm win}$26

derived using a functional delta method (Mutis et al., 1 May 2026). Computation uses a hybrid IRLS-Newton algorithm with IRLS updates for $t_{\rm win}$27, Newton-Raphson for $t_{\rm win}$28, and a fixed-point Fisher-consistent scale update for $t_{\rm win}$29 (Mutis et al., 1 May 2026). The paper states that this yields “stable, reproducible convergence” and has been implemented in the fcsar R package (Mutis et al., 1 May 2026).

6. Related but distinct usage: multiscale Fisher’s independence test

A separate line of work uses “M-FISHER” or MultiFIT for a nonparametric independence test rather than a representation model, adaptation method, or Fisher-information design framework (Gorsky et al., 2018). The method tests independence between random vectors by transforming marginals to $t_{\rm win}$30, constructing canonical nested dyadic partitions,

$t_{\rm win}$31

and forming $t_{\rm win}$32 contingency tables on adaptively selected cuboids (Gorsky et al., 2018).

The key equivalence result states that $t_{\rm win}$33 if and only if for every cuboid $t_{\rm win}$34 and pair $t_{\rm win}$35 the local odds ratio

$t_{\rm win}$36

equals 1 (Gorsky et al., 2018). Each local null is then tested by Fisher’s exact test on the table

$t_{\rm win}$37

with a closed-form exact $t_{\rm win}$38-value and “No permutation” required (Gorsky et al., 2018).

The method proceeds in a coarse-to-fine adaptive scan, expanding only those cuboids whose local Fisher tests satisfy $t_{\rm win}$39, followed by FWER control using either a holistic or resolution-specific correction scheme (Gorsky et al., 2018). A central theoretical claim is finite-sample validity under adaptive selection: even when $t_{\rm win}$40 is selected adaptively, the conditional distribution of the selected table count remains hypergeometric under independence (Gorsky et al., 2018). The source states strong FWER control without resampling or asymptotic approximation and gives expected computational complexity $t_{\rm win}$41 under typical null settings, contrasting it with $t_{\rm win}$42 or worse for distance-covariance, HHG, and dHSIC (Gorsky et al., 2018). It also reports practical scalability up to $t_{\rm win}$43 in simulations and a flow-cytometry application with $t_{\rm win}$44 cells that ran in about 5 minutes on a laptop (Gorsky et al., 2018).

This usage is related in name only. The “Fisher” in MultiFIT refers to Fisher’s exact test, not Fisher information, natural gradient geometry, or Fisher-consistent M-estimation.

7. Conceptual relations and recurrent misconceptions

Despite the shared label, the M-FISHER variants are connected only at a high level by their reliance on distinct “Fisher” constructs. In the industrial foundation model, “FISHER” is a model name and the technical core is STFT sub-band modeling plus teacher-student self-distillation (Fan et al., 22 Jul 2025). In martingale-driven prompt adaptation, the essential ingredients are non-conformity scores, e-processes, Ville’s inequality, and Fisher-preconditioned updates (Khan et al., 4 Oct 2025). In optical metrology, the core object is the Fisher-information matrix derived from scattering-matrix derivatives (Verreussel et al., 31 May 2026). In spatial regression, the defining property is Fisher consistency of a redescending M-estimator under a bias-corrected SAR formulation (Mutis et al., 1 May 2026). In MultiFIT, the “Fisher” component is Fisher’s exact test on adaptively refined $t_{\rm win}$45 tables (Gorsky et al., 2018).

One common misconception is to assume that all M-FISHER methods are variants of the same algorithmic family. The source material does not support that interpretation. Another is to treat “Fisher” as always meaning Fisher information. That is incorrect for MultiFIT, and only partly correct for the prompt-adaptation and optical-metrology variants. A third misconception is to regard the industrial M-FISHER as the canonical or sole meaning of the term. The details explicitly indicate that this name was assigned to a paper whose published model name is simply FISHER (Fan et al., 22 Jul 2025).

This suggests that “M-FISHER” functions best as a context-dependent label rather than a stable technical term. In practice, the intended meaning is usually recoverable from neighboring vocabulary: SCADA, STFT sub-bands, and RMIS indicate the industrial foundation model; e-process, Ville’s inequality, and Fisher prompting indicate sequential adaptation; scattering matrix, D-optimality, and CRLB indicate optical metrology; Fisher-consistent redescending M-estimation and spatial scalar-on-function regression indicate robust spatial statistics; and dyadic partitions together with Fisher’s exact test indicate MultiFIT.