Frequency Selection: Methods and Applications

• Frequency Selection is a technique used to choose, rank, and optimize frequency bands or components for efficient signal processing, denoising, and system control.
• Methods vary from adaptive early exits in spectrogram analysis to threshold-based top-K retention and dynamic hyperparameter tuning based on modal frequency.
• Applications span audio and image processing, power system optimization, cognitive radio, and modeling frequency-dependent selection in biological and cultural contexts.

Frequency selection denotes a family of procedures that choose, rank, gate, or optimize over frequencies, frequency bands, or frequency-conditioned interactions. In contemporary research, the term appears in at least two distinct but mathematically related senses. In signal-processing and machine-learning settings, it usually means selecting spectral components, sub-bands, or modal frequencies to improve efficiency, denoising, reconstruction, or control. In evolutionary and cultural dynamics, it denotes selection whose strength depends on current population frequency rather than on trait identity alone. Across these uses, frequency selection operates as a mechanism for reducing a large hypothesis space to the components that materially affect prediction, inference, or dynamics (Wang et al., 2021, Wu et al., 1 Aug 2025, Mao et al., 2021, Sehloff et al., 2021, Newberry et al., 2021, Schraiber et al., 2016).

1. Core formulations

In spectral settings, frequency selection is typically posed over a transformed representation. A representative example is the FreK module in KFS, where a real-valued time series is mapped by FFT, bin energies are computed as $E_k = |X_k|^2$, and the smallest $K$ is chosen such that the cumulative energy ratio exceeds a threshold $\delta$; the selected spectrum is then reconstructed by inverse FFT (Wu et al., 1 Aug 2025). Other formulations replace energy ranking with sequential computation. HIDACT partitions a log-Mel spectrogram into contiguous sub-bands and processes them from low to high frequency, maintaining an uncertainty-controlled accumulated output and halting when the lower bound for the current top class exceeds the upper bound for the runner-up, namely $l[c^*] = a_t[c^*](1-p)^{K-k}$ and $u[c'] = a_t[c'] + p \cdot (K-k)$, with early exit declared when $l[c^*] > u[c']$ (Wang et al., 2021).

A second formulation treats frequency as an operating variable rather than as a spectral coordinate. In embedded LFAC transmission, the subnetwork frequency $\omega_l$ is a continuous OPF decision variable, and it directly changes series reactance and shunt susceptance through $X = \omega L$ and $B^{\mathrm{sh}} = \omega C$, thereby moving the system between angle-constrained, thermal-constrained, and voltage-constrained regimes (Sehloff et al., 2021). In game optimization, modal frequency is inferred from Jacobian eigenvalues: if $T_i = 1 - \gamma \lambda_i = R_i e^{-i\theta_i}$, then the angle $K$0 is the per-step discrete rotation frequency used by Modal LookAhead to choose stabilizing hyperparameters (Sanyal et al., 26 Jan 2026).

In evolutionary work, the same phrase shifts meaning. Here the relevant object is not a Fourier axis but a selection function of prevalence. In cultural evolution, frequency-dependent selection is described by a selection coefficient $K$1, where $K$2 is current trait frequency; positive FDS favors common traits, while negative FDS favors rare traits (Newberry et al., 2021). In Wright–Fisher diffusion models with linear frequency dependence, the drift can take the form $K$3, so the sign and magnitude of selection change with allele frequency itself (Pfaffelhuber et al., 2012). This corpus therefore uses “frequency selection” both for choosing frequencies and for selection driven by frequency.

2. Adaptive selection in audio and time-series models

In sound event detection, HIDACT operationalizes adaptive frequency selection as monotonic processing of contiguous spectrogram bands. The input is a log-Mel representation $K$4, partitioned into $K$5 non-overlapping sub-bands; in the reported experiments, $K$6 and $K$7, so each sub-band contains 16 Mel bins. HIDACT builds on a Multi-View Network unrolled across frequency and time, trains with all bands using Differentiable Adaptive Computation Time, and applies hard halting only at test time (Wang et al., 2021). The empirical result is that HIDACT and ACT achieve roughly 70% macro F1 within an average of 5 computational steps per frame, while HIDACT processes fewer than 6 sub-bands on average and attains the highest F1 among sub-band models. The same study reports that performance is comparable to FB3 with much fewer parameters and roughly one-third of the MACs, and the deployment discussion states that HIDACT delivers comparable accuracy to larger full-band CRNNs with one-third MACs and two orders of magnitude fewer parameters.

In long-term forecasting, KFS uses frequency selection primarily as denoising and dominant-harmonic isolation. For each scale and channel, FreK computes an FFT, ranks bins by power, keeps the smallest set covering a prescribed cumulative energy level, and reconstructs a filtered signal by inverse FFT. The paper’s ablation reports that error metrics improve as $K$8 increases and reach minima near $K$9 on ETTh1 and ETTh2 with look-back 96 and forecast horizon 96 (Wu et al., 1 Aug 2025). Its filter study gives concrete pairwise comparisons: on Weather, Top-K achieves average MSE/MAE of 0.243/0.271 versus 0.248/0.274 for the average filter and 0.245/0.272 for the Gaussian filter; on ETTh2, Top-K gives 0.367/0.394 versus 0.383/0.405 and 0.372/0.397. In this setting, frequency selection is not adaptive halting but thresholded retention of dominant spectral energy.

These two systems illustrate a recurring distinction. HIDACT selects as little frequency information as needed for a confident decision on each frame, whereas FreK retains enough frequency information to preserve approximately 90% cumulative energy at each scale. This suggests two major design families: compute-adaptive frequency selection, where the objective is early stopping, and representation-adaptive frequency selection, where the objective is denoising or harmonic concentration.

3. Imaging, reconstruction, and steganographic embedding

In image deblurring, frequency selection is implemented as a nonlinear operator in the Fourier domain. The defining operation is $\delta$0, where ReLU is applied separately to the real and imaginary parts of the spectrum. The paper argues that this simple clipping exposes blur direction, level, and kernel-like structure after inverse transform, and incorporates it into a Res FFT-ReLU Block that combines a spatial residual stream with a frequency stream (Mao et al., 2021). When inserted into NAFNet, the method reaches 33.85 dB PSNR on GoPro. Reported plug-in gains include DeepDeblur 31.15 $\delta$1 32.37, U-Net 29.20 $\delta$2 30.39, MPRNet-small 31.09 $\delta$3 32.50, MIMO-UNet 31.90 $\delta$4 32.71, and NAFNet32 32.95 $\delta$5 33.12.

In active view selection for Gaussian Splatting, frequency selection is used to score candidate images. The method renders candidate views with the current GS model, computes their spectra, and selects the view with the lowest median frequency, because low-frequency dominance indicates blur or under-constrained geometry (Li et al., 2024). Experiments on Playroom, Dr. Johnson, Truck, and Train start from 10 images and select up to 100 views. The paper reports trajectory length reductions to about 25–30% of the original dataset path length, for example Train 332.39 $\delta$6 60.16 and Truck 406.72 $\delta$7 104.74, while rendering quality remains close to the full-dataset baseline on Train, Truck, and Playroom.

FAST generalizes this idea from per-image selection to dataset compression. It matches empirical characteristic functions in the frequency domain and identifies a “vanishing phase gradient” pathology at medium and high frequencies. Its Attenuated Phase-Decoupled CFD adds an explicit phase penalty, $\delta$8, with $\delta$9, and its Progressive Discrepancy-Aware Sampling schedules frequencies from low to high norm (Cui et al., 22 Nov 2025). The reported experimental summary states an average accuracy gain of 9.12%, a 96.57% reduction in power consumption, and a 2.2x average speedup.

In color image steganography, frequency selection is tied to perceptual masking. The cited model selects the RGB B channel as the embedding channel and the DWT diagonal high-frequency sub-band $l[c^*] = a_t[c^*](1-p)^{K-k}$0 as the embedding domain, on the grounds that the human visual system is less sensitive to blue perturbations and to high-frequency detail (Su et al., 2021). With this choice, the reported metrics are C_Error 0.66, C_PSNR/CL-PSNR 82.31/44.33, C_SSIM 0.9975, S_PSNR 37.75, and S_SSIM 0.9999 at 8 bpp. The ablation against low-frequency-only embedding shows the opposite regime: using $l[c^*] = a_t[c^*](1-p)^{K-k}$1 produces C_Error 5.51, C_PSNR/CL-PSNR 76.17/27.29, and C_SSIM 0.9211.

4. Control, optimization, and physical system design

In power systems, frequency selection becomes a continuous operating decision. The LFAC transmission study formulates AC OPF with subnetwork frequencies $l[c^*] = a_t[c^*](1-p)^{K-k}$2 as variables, together with generator dispatch, MMC interface injections, voltages, and angles (Sehloff et al., 2021). Frequency reduction lowers reactance, changes shunt susceptance, and can move a corridor from an angle-constrained region to a thermal plateau and then to a voltage-constrained region. In the Nordic case, the paper reports that optimal $l[c^*] = a_t[c^*](1-p)^{K-k}$3 typically lies in the 19–36 Hz range for single or few boundary-line upgrades, while large multi-terminal corridors can prefer very low frequencies around 2–6 Hz. The comparison with HVDC finds that LFAC-OPF matches or beats the best HVDC configuration across the examined point-to-point upgrades.

In underlay MIMO cognitive radio, the selection variable is a frequency band rather than a continuous frequency. The SU chooses one band per slot and controls transmit power under an interference leakage constraint. A key formula is $l[c^*] = a_t[c^*](1-p)^{K-k}$4, so stale null-space information directly raises leakage (Chaudhari et al., 2019). The paper proves that dynamic band-hopping policies such as random, round robin, and DSEE produce higher interference than fixed-band policies, because the effective null-space age $l[c^*] = a_t[c^*](1-p)^{K-k}$5 is larger under hopping. It also shows that fixed-band dynamic power control yields no smaller expected rate than fixed-band fixed power control, and that the rate gap to a clairvoyant policy shrinks as temporal correlation and the number of SU antennas increase.

In smooth games, frequency selection is neither spectral denoising nor resource allocation but hyperparameter tuning through oscillatory mode estimation. Modal LookAhead extracts the dominant modal frequency from the Jacobian spectrum and then chooses the LookAhead depth $l[c^*] = a_t[c^*](1-p)^{K-k}$6 and interpolation coefficient $l[c^*] = a_t[c^*](1-p)^{K-k}$7 so that the inner iterations accumulate approximately a $l[c^*] = a_t[c^*](1-p)^{K-k}$8 phase shift on that mode while respecting the stability cap $l[c^*] = a_t[c^*](1-p)^{K-k}$9 (Sanyal et al., 26 Jan 2026). The paper proves convergence for monotone Lipschitz variational inequalities and reports faster convergence than GD, EG, OGD, random LA, Adam, and LA–Adam in bilinear and strongly convex–strongly concave regimes.

Quantum machine learning pushes the same idea into feature-map design. Angle encoding generates truncated Fourier series, but mixed frequencies can grow exponentially with repetition depth and dimensionality. The proposed remedy is explicit frequency selection, $u[c'] = a_t[c'] + p \cdot (K-k)$0, together with dimensional separation that allows mixed terms only within known interacting groups (Poppel et al., 14 Aug 2025). In the four-dimensional example, restricting mixing to two separable two-dimensional blocks reduces cardinality from $u[c'] = a_t[c'] + p \cdot (K-k)$1 to $u[c'] = a_t[c'] + p \cdot (K-k)$2. The paper reports successful noisy-simulator training and hardware inference on IBM Fez, with hardware $u[c'] = a_t[c'] + p \cdot (K-k)$3 in the 2D case and $u[c'] = a_t[c'] + p \cdot (K-k)$4 in the 4D case.

5. Feature, model, and representation selection

A distinct line of work uses frequency selection as a feature-ranking principle. In brain-tumour spectroscopy, the 512-point $u[c'] = a_t[c'] + p \cdot (K-k)$5H-MRS axis is divided into an artifact-dominated zone $u[c'] = a_t[c'] + p \cdot (K-k)$6, a metabolite-rich central band $u[c'] = a_t[c'] + p \cdot (K-k)$7, and a noise-dominated tail $u[c'] = a_t[c'] + p \cdot (K-k)$8 (Arizmendi et al., 11 Mar 2025). The selection score is a Moving Window ratio $u[c'] = a_t[c'] + p \cdot (K-k)$9, and the paper reports that $l[c^*] > u[c']$0 yields the largest discrimination scores. Zone energies are normalized by the noise-zone energy $l[c^*] > u[c']$1, and top-ranked frequencies are added in approximately 1% cumulative-energy groups. Example results include G1 versus normal with mean 98.89% $l[c^*] > u[c']$2 2.5 and best 100.0% $l[c^*] > u[c']$3 0.0, and gl versus me with mean 68.38% $l[c^*] > u[c']$4 9.0 and best 75.69% $l[c^*] > u[c']$5 9.8.

Text categorization uses an analogous but non-spectral frequency statistic. The term-frequency t-test computes

$l[c^*] > u[c']$6

with $l[c^*] > u[c']$7 and pooled within-class variance $l[c^*] > u[c']$8 (Wang et al., 2013). The method is designed to address two shortcomings of document-frequency-based feature selection: unreliability for low-frequency terms and the neglect of within-document term counts. The reported experiments on Reuters-21578 and 20 Newsgroups show that the t-test is comparable to or slightly better than $l[c^*] > u[c']$9 and IG in macro-$\omega_l$0 and micro-$\omega_l$1.

Model selection itself can be frequency-based. Prediction Weighted Maximum Frequency Selection records how often each sparse model appears at a given dimension across bootstrap-resampled solution paths, defines the maximum-frequency model

$\omega_l$2

and then multiplies these frequencies by prediction-based weights to form WMF (Liu et al., 2017). The procedure is developed for adaptive LASSO and adaptive Elastic-Net with diverging $\omega_l$3, and the paper proves consistent model selection with favorable convergence rates.

Static word embedding introduces yet another meaning of frequency-aware selection. The MPD criterion chooses embedding dimension without training embeddings by comparing oracle matrices from two corpus partitions and combining a primitive Gram-matrix distance with a post-processed distance that suppresses frequency-dominated principal directions (Shen et al., 2023). The reported efficiency–performance comparison gives average task performance 59.1 for MPD, versus 57.8 for PIP and 57.3 for PCA, while grid-search baselines are 14.01x, 28.02x, and 56.04x slower depending on granularity.

6. Frequency-dependent selection in culture and population genetics

In cultural evolution, frequency selection is explicitly frequency-dependent selection. The inferred function $\omega_l$4 links trait frequency to growth, so neutrality corresponds to $\omega_l$5, positive FDS favors common traits, and negative FDS favors rare traits (Newberry et al., 2021). The paper’s central empirical result is strong negative FDS in baby names: the most common names tend to decline by 2%–6% per year, whereas rare names around 1 in 10,000 births tend to increase by 1%–3% per year. In the United States, the point estimates are +1.4% per year at $\omega_l$6 and $\omega_l$7 per year at $\omega_l$8. The same study also infers a fixed fitness offset of roughly 0.66% per year disfavoring extant female names relative to male names, and a strong advantage for biblical names across all frequency classes.

In population genetics, frequency-dependent selection is formalized in diffusion and genealogical models. For weak selection in a finite population, the diffusion drift can be written as $\omega_l$9, and the fixation probability of a single mutant satisfies

$X = \omega L$0

which yields the one-third law condition $X = \omega L$1 (Pfaffelhuber et al., 2012). Later work extends this perspective using killed and pruned lookdown ancestral selection graphs for Moran models with mutation and frequency-dependent selection, and by constructing $X = \omega L$2-Fleming–Viot jump-diffusion limits dual to branching–coalescing ancestral processes with general branching and simultaneous multiple collisions (Baake et al., 2020, Casanova et al., 2016). These results make present-type and ancestral-type distributions computable from dual genealogical processes rather than from the forward process alone.

Ancient-DNA time-series inference returns to direct frequency observations. A Bayesian method for allele frequency time series models the derived-allele trajectory by Wright–Fisher diffusion with general diploid selection and time-varying demography, augments the latent path in MCMC, and estimates selection coefficients together with allele age (Schraiber et al., 2016). In the horse coat-color application, incorporating demographic history materially changes the inferred mode and strength of selection for both MC1R and ASIP; the study states that ignoring a relevant demographic history can significantly bias the results of inference. Across these biological and cultural settings, frequency selection is not a choice over frequencies but a statement that the force of selection itself varies with how common a type already is.

Taken together, these literatures show that frequency selection is not a single method but a reusable abstraction. It can mean early exit over spectrogram bands, Top-K retention of energetic Fourier bins, phase-aware low-to-high curricula, operating-frequency optimization in infrastructure, mode-aware hyperparameter tuning in games, selective retention of Fourier support in quantum models, or frequency-dependent fitness in stochastic evolution. What unifies these cases is the same structural move: replacing undifferentiated treatment of all components with a rule that privileges the frequencies, bands, or frequency classes that most strongly govern the phenomenon of interest.