MM-Detect: Signal Recovery in Diverse Domains

• MM-Detect is a multifaceted detection framework that extracts weak, structured signals from noisy data across varied applications.
• It employs tailored strategies like symmetry metrics in gerrymandering, score-based diffusion in radar, and advanced detector hardware in astrophysics.
• Across domains, MM-Detect emphasizes preserving key observables and mitigating background interference to ensure robust signal inference.

Searching arXiv for “MM-Detect” and closely related terms to verify whether it is a standardized term or a context-dependent label. MM-Detect, as the cited arXiv literature suggests, is not a single standardized apparatus or algorithm but a context-dependent designation spanning several technically distinct detection regimes. In one usage, “MM” denotes the Mean-Median Difference, a partisan-symmetry statistic for gerrymandering analysis; in others it refers to millimeter/submillimeter detection, micro-TPC matrix detection, or model- and data-driven multitarget inference under mainlobe jamming. Across these settings, the common technical problem is the recovery of weak structured signals in the presence of nontrivial backgrounds, whether those observables are district vote-share distributions, radar returns, nuclear-recoil tracks, or mm/submm fluxes and spectra (Deford et al., 2024, Guo et al., 27 Nov 2025, Santos et al., 2011, Yoo et al., 2017).

1. Terminological scope and domain-specific meanings

The literature associates MM-Detect with several non-equivalent detection concepts. One branch centers on the Mean-Median Difference, defined on district vote shares and used as a symmetry metric in partisan-gerrymandering analysis. Another concerns radar detection under mainlobe jamming, where a diffusion-based Model and Data Dual-driven framework combines learned jamming priors with sparse Bayesian learning. A third branch concerns detector hardware and signal extraction in particle and astrophysical instrumentation, including the MIMAC micro-TPC matrix for directional dark-matter searches, the JCMT Transient Survey’s sub-mm variability detection, and integrated mm/submm detector architectures such as SuperSpec and cold-electron bolometers (Deford et al., 2024, Guo et al., 27 Nov 2025, Santos et al., 2010, Shirokoff et al., 2012).

This heterogeneity is itself informative. It suggests that MM-Detect is best understood operationally: a family resemblance among methods that infer rare or structured phenomena from noisy measurements, rather than a single canonical formalism. The cited works converge on a recurring architecture of inference: explicit signal models, careful treatment of backgrounds, and detector-specific observables that preserve directional, spectral, temporal, or morphological information.

2. Mean-Median Difference as a detection metric for partisan asymmetry

In the gerrymandering literature, the Mean-Median Difference, or MM, is defined from district vote shares $V_1\le \cdots \le V_n$ for party $A$ as

$$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$

It is interpreted through the seats-votes curve as the horizontal displacement from $(0.5,0.5)$ to the point where the curve intersects $S=0.5$. A “perfect” score is $MM=0$. The related Partisan Bias metric is

$$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$

with $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$, and is the corresponding vertical displacement at $V=0.5$ (Deford et al., 2024).

The theoretical analysis in “Bounds and Bugs: The Limits of Symmetry Metrics to Detect Partisan Gerrymandering” establishes sharp range results for MM and PB under equal turnout. The standing feasibility constraints are

$$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$

A central conclusion is that the region of $A$0 for which $A$1 is exactly the same as the region for which $A$2 when turnout is equal. Under unequal turnout with maximum-to-minimum turnout ratio $A$3, the admissible zero-set expands substantially, and as $A$4 the region of $A$5 with $A$6 or $A$7 expands to fill the whole unit square $A$8 (Deford et al., 2024).

The empirical and conceptual criticism is that MM does not reliably become more extreme when a party wins more districts. The paper’s short-burst experiments show that as the number of Democratic-won districts increases, MM does not systematically increase, and as Republican-won districts increase, MM does not systematically decrease. On 18 different U.S. maps, MM and PB can fail to detect extreme seats outcomes; the largest observed $A$9 difference in the data is $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$0, on South Carolina’s congressional map. Specific examples include Massachusetts congressional maps, where both MM and PB suggest that the map may be a Republican gerrymander, and Pennsylvania’s 2011 congressional map, where PB fails to flag a map widely recognized as gerrymandered. In this usage, MM-Detect is therefore a problematic label if the intended target is partisan advantage measured through districts won rather than symmetry alone (Deford et al., 2024).

3. Model- and data-driven multitarget detection under mainlobe jamming

In radar, MM-Detect is represented by the diffusion-based Model and Data Dual-driven approach, DMDD, for multitarget detection under structured mainlobe jamming. The pulse-compressed measurement model is

$$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$1

which is rewritten on a discretized range grid as

$$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$2

with $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$3 sparse and $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$4. The jamming distribution $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$5 is learned from target-free data, and its score $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$6 is modeled by a score-based diffusion process (Guo et al., 27 Nov 2025).

The forward diffusion is given by

$$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$7

and the learned score network $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$8 is trained by denoising score matching to approximate $$MM=\operatorname{median}\{V_1,\dots,V_n\}-\operatorname{mean}\{V_1,\dots,V_n\}.$$9. The posterior score driving reverse-time sampling is approximated by

$(0.5,0.5)$0

Target amplitudes are modeled with a sparse Bayesian learning prior,

$(0.5,0.5)$1

and the hyperparameters and noise variance are updated by EM. Only one diffusion process is constructed, for jamming; target states are estimated through direct posterior inference, which the paper identifies as the main efficiency advantage over approaches that diffuse both signal and interference (Guo et al., 27 Nov 2025).

The numerical results are specific. In the synthetic on-grid case, pulse compression fails, DM-SBL detects both targets but still has many false alarms, and DMDD detects both targets with only one false alarm. Runtime is reported as $(0.5,0.5)$2 s for DM-SBL and $(0.5,0.5)$3 s for DMDD. Under $(0.5,0.5)$4 dB and integrated SNR from 12 to 22 dB, DMDD reaches $(0.5,0.5)$5 up to $(0.5,0.5)$6 with false alarm probability about $(0.5,0.5)$7, compared with around $(0.5,0.5)$8 for ADMM and DM-SBL and around $(0.5,0.5)$9 for SBL and SBL-SOM. At $S=0.5$0, DMDD requires integrated SNR $S=0.5$1 dB, whereas DM-SBL requires integrated SNR $S=0.5$2 dB, yielding about $S=0.5$3 dB gain. On real clutter data with injected targets, DMDD achieves lower false-alarm probability and about $S=0.5$4 dB performance gain over DM-SBL (Guo et al., 27 Nov 2025).

4. Directional particle detection and machine-detector background control

A hardware-centric MM-Detect paradigm appears in MIMAC, the MIcro-tpc MAtrix of Chambers for directional dark-matter detection. Its physical premise is that WIMP recoils should be anisotropic because the Solar System moves through the Galactic halo at about $S=0.5$5, so the laboratory frame should see an incoming WIMP wind preferentially from the direction of the Cygnus constellation. MIMAC is a gaseous micro-TPC matrix using light nuclei such as $S=0.5$6He, $S=0.5$7, $S=0.5$8, H, and $S=0.5$9F, with a particular emphasis on axial, spin-dependent interactions on odd nuclei (Santos et al., 2010, Santos et al., 2011).

The detector combines a drift volume, bulk micromegas amplification, a segmented anode, and fast self-triggered electronics. The micromegas amplification gap is $MM=0$0 or $MM=0$1; one prototype has a $MM=0$2 active readout area sampled every $MM=0$3 ns, and another has a $MM=0$4 active area with $MM=0$5 pitch sampled at $MM=0$6 MHz, i.e. every $MM=0$7 ns, using 8 ASIC chips with 64 channels each for 512 channels total. The $MM=0$8 coordinate is reconstructed from drift time according to

$MM=0$9

Because pure $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$0 drifts too quickly, the mixture $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$1 is used; adding about $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$2–$$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$3 slows the drift sufficiently that recoils in the $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$4–$$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$5 keV range produce at least $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$6–$$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$7 time slices (Santos et al., 2011).

Background rejection is based on energy and track morphology. For a given ionization energy, an electron track is about an order of magnitude longer than a nuclear recoil track. MIMAC also introduces the normalized integrated straggling, NIS, defined as the sum of all angular deviations along the reconstructed track normalized by the ionization energy. A $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$8F recoil with 5 keV kinetic energy produces only 1.2 keV of ionization, making ionization-quenching measurements essential. The reported prototypes reconstruct a 5.9 keV electron track, an 8 keV proton recoil with 2.4 mm track length, a $$PB=\frac{1}{2}\left(\text{Proportion of }V_i\text{ larger than }\overline V-\text{Proportion of }V_i\text{ smaller than }\overline V\right),$$9F recoil track around 3 mm, a 57 keV proton track in $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$0 at 50 mbar, and a 40 keVee fluorine track. The scale-up path runs from a bi-chamber prototype to a $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$1 demonstrator and eventually a $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$2 matrix detector; for the $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$3 demonstrator, the paper cites discovery down to about $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$4 pb and exclusion down to about $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$5 pb in the spin-dependent proton cross-section plane (Santos et al., 2010, Santos et al., 2011).

A distinct but related detector-background problem appears in the muon collider machine-detector interface. There the dominant source is muon decays: at $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$6 TeV beam energy with $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$7 muons per bunch, the decay rate is $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$8 and $\overline V=\operatorname{mean}\{V_1,\dots,V_n\}$9 for two beams. The resulting radiation power in the ring is approximately $V=0.5$0. Mitigation relies on MARS15-based optimization of masks, nozzles, shielding, and magnet geometry, especially an open-midplane dipole with 160 mm coil aperture, 55 mm gap height, 6 m magnetic length, and 8 T nominal field. The integrated shielding strategy reduces backgrounds by more than three orders of magnitude, and realistic detector simulations indicate that backgrounds are manageable with timing better than 0.5 ns and a 2–3 ns time gate (Mokhov, 2012).

5. Millimeter and submillimeter source detection in astronomy

In astronomy, MM-Detect is closely associated with millimeter and submillimeter detection of compact, transient, or deeply embedded sources. The JCMT Transient Survey provides a prototypical case. Serpens Main was monitored with repeat, well-calibrated $V=0.5$1 SCUBA-2 maps taken roughly monthly, within a program covering eight nearby star-forming regions at 450 and $V=0.5$2. The Serpens Main data comprise 12 observations between 2 February 2016 and 17 April 2017, with 11 epochs used for the final $V=0.5$3 analysis. Relative flux calibration using five stable calibrator clumps reduced the effective flux uncertainty to about 2–3%, enabling the detection that Source 5, coincident with the Class I protostar EC 53, brightened while nearly all other clumps in the field remained stable (Yoo et al., 2017).

The EC 53 light curve shows a quiescent phase from February to August 2016 at about $V=0.5$4 with standard deviation $V=0.5$5, brightening beginning in September 2016, an estimated peak near $V=0.5$6, and slow fading from February 2017 to April 2017. The average brightness in the last four epochs was $V=0.5$7, quoted as $V=0.5$8 above the first-six-epoch mean; the most enhanced point was $V=0.5$9 above the faint level. The $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$0 increase in $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$1 flux is interpreted as dust heating in the envelope, generated by a protostellar luminosity increase of at least a factor of 4, with $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$2 and $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$3 in a blackbody-like approximation. The $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$4 light curve resembles the historical $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$5-band light curve, which varies by a factor of $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$6 with a 543-day period and is interpreted as accretion variability excited by interactions between the accretion disk and a close binary system (Yoo et al., 2017).

Pre-ALMA GRB studies show a different mm/submm detection regime. The compiled catalogue contains 102 GRBs, including 88 afterglow searches and 22 afterglow detections, with redshifts from 0.168 to 8.2. The synchrotron peak often passes through the mm/submm band within hours to days, with

$$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$7

The overall afterglow detection fraction is $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$8, attributed explicitly to instrument sensitivity rather than absence of mm/submm emission. Band-by-band afterglow detections are 18/61 in Band 3, 1/2 in Band 4, 6/38 in Band 6, 5/31 in Band 7, and 0/6 in Band 9. The paper reports a correlation between X-ray flux density at 0.5 days and mm/submm peak flux density, and estimates that full ALMA could detect up to 98% of afterglows if observed near the synchrotron peak (Postigo et al., 2011).

L1451-mm illustrates the continuum-detection problem at the earliest stages of low-mass star formation. CARMA detects a compact 3-mm source with a point-source fit of $$0\le V\le 1,\qquad 0\le S\le 1,\qquad S\le 2V,\qquad S\ge 2V-1.$$9 mJy, while SMA detects a compact 1.3-mm continuum component with flux 27.0 mJy and a compact mass $A$00 for $A$01 K. Spitzer shows no point source in the mid-IR, implying $A$02 and $A$03 K, and the SMA $A$04 data reveal a slow, poorly collimated outflow with characteristic velocity $A$05 and dynamical time $A$06 yr. Simultaneous fitting of the broadband SED and interferometric visibilities requires a central heating source but cannot discriminate decisively between a dense core with a YSO and disk and a dense core with a first hydrostatic core (Pineda et al., 2011).

6. Detector architectures for mm/submm spectroscopy and surveys

SuperSpec exemplifies an integrated on-chip mm-wave detector architecture. It consists of a horn-coupled microstrip feedline, a bank of narrow-band superconducting resonator filters, and TiN kinetic inductance detectors that absorb the admitted power. The intended signal path is: incoming mm-wave radiation, feedline transport, resonant filter selection of a narrow band, TiN absorber conversion of power to quasiparticles, and readout via KID resonance shift. The prototype mm-wave microstrip feedline and spectral filters are designed for 195–310 GHz and fabricated from niobium with $A$07 K, while the KIDs are designed to operate at hundreds of MHz and are fabricated from titanium nitride with $A$08 K (Shirokoff et al., 2012).

The mm-wave channel resolving power is

$A$09

with maximum absorption at $A$10. For 600 channels, $A$11, over 195–310 GHz, the oversampling factor is $A$12, giving about 80% in-band absorption efficiency. The dark prototype includes 74 tuned mm-wave filters spanning 200–300 GHz. In initial laboratory testing, 74 of 77 typical KIDs on the main feedline and 3 of 4 low-frequency termination KIDs were observed; all measured resonances were shifted downward in frequency by about 55% relative to design, and the measured coupling quality factors were tightly clustered with mean $A$13. The stated demonstration instrument consists of two 500-channel, $A$14 spectrometers, one in the 1-mm atmospheric window and the other covering the 650 and 850 micron bands (Shirokoff et al., 2012).

Cold-electron bolometers represent a different mm/submm survey technology, motivated by future balloon-borne and spaceborne observations requiring detectors insensitive to cosmic rays and with

$A$15

Their physical basis is a very small absorber volume and thermal decoupling of absorber electrons from the phonon system. The sample developed for the 350 GHz channel of OLIMPO contains 144 CEBs connected in series and parallel, setting the saturation limit to about 41 pW; the measured bandwidth is 29% wide centered at 375 GHz, and the sample is cooled to 305 mK using a two-stage $A$16He fridge (Salatino et al., 2014).

Optical tests using chopped 77–300 K blackbody radiation produced, in one configuration, optical responsivity $A$17, measured white noise about $A$18, and output signal about $A$19; in another configuration, the optical responsivity was $A$20 and the optical NEP was $A$21. X-ray tests using a microfocus source were used as a proxy for cosmic-ray sensitivity. The paper concludes that substrate-to-absorber coupling is weak and that the detector is therefore insensitive to high-energy particle hits, a property regarded as important for OLIMPO and LSPE integration (Salatino et al., 2014).

7. Common methodological patterns, limitations, and interpretive significance

Taken together, these works suggest that MM-Detect problems are organized less by disciplinary boundary than by a shared inference structure. Signal recovery depends on preserving physically meaningful observables: district-vote distributions for MM and PB, recoil direction and straggling for MIMAC, predictor-corrector posterior sampling and sparse priors for DMDD, relative flux calibration for EC 53, and superconducting resonance shifts or SIN-junction thermal balance for mm/submm detector hardware (Deford et al., 2024, Santos et al., 2011, Guo et al., 27 Nov 2025, Yoo et al., 2017, Shirokoff et al., 2012, Salatino et al., 2014).

The same literature also emphasizes that detectability is often limited by the mismatch between the measured observable and the latent phenomenon of interest. In gerrymandering, symmetry metrics can be zero over a broad range of seats-votes outcomes and therefore need not track districts won. In protostellar monitoring, the 850 $A$22 response is a diluted thermal echo because the observed emission blends varying and non-varying envelope regions and the outer envelope is also heated by the interstellar radiation field. In pre-ALMA GRB work, the low detection fraction is attributed to instrument sensitivity, not to absence of afterglow emission. In muon-collider detector design, raw backgrounds are so large that feasibility depends on integrated shielding and fast timing rather than on detector response alone (Deford et al., 2024, Yoo et al., 2017, Postigo et al., 2011, Mokhov, 2012).

A plausible implication is that MM-Detect, across its disparate usages, is most informative when its “MM” component is treated not as a nominal label but as a statement about which structure is being preserved under inference: symmetry in electoral data, directionality in recoil maps, morphology in tracks, learned priors in jamming, or spectral selectivity in mm/submm instrumentation. The cited literature is therefore unified by a technical principle rather than a shared ontology: reliable detection requires observables whose distortion under background, calibration error, or model misspecification remains sufficiently constrained to support inference.