Oversampling: Concepts and Applications

• Oversampling is the intentional use of sampling rates or synthetic data beyond the minimal required resolution to recover lost information in various applications.
• In communications, oversampling boosts performance by increasing temporal resolution, thereby compensating for quantization losses and channel uncertainties.
• In machine learning, techniques like SMOTE generate synthetic samples to rebalance imbalanced datasets, enhancing classifier performance and decision region geometry.

Searching arXiv for relevant papers on oversampling across communications, machine learning, signal processing, and numerical methods. Oversampling denotes the deliberate use of a sampling rate, representation density, or synthetic sample count that exceeds a baseline resolution regarded as minimal for a given task. In signal processing and communications, the baseline is typically the Nyquist or symbol rate, and oversampling means sampling at a factor $\eta>1$ or $M>1$ above that rate. In machine learning, especially under class imbalance, oversampling denotes enlargement of the minority class beyond its empirical frequency, either by replication or by synthesizing new samples. In numerical multiscale methods, oversampling denotes the use of local computational patches larger than a target coarse element. Across these settings, the common purpose is to recover information otherwise lost to coarse discretization, severe quantization, model uncertainty, or scale truncation. The term therefore refers not to a single technique but to a family of strategies that trade increased sample density, higher-dimensional local context, or synthetic data augmentation for improved inference, estimation, detection, or approximation performance (Koch et al., 2010).

1. Conceptual scope and formal definitions

In continuous-time communication models, oversampling is defined relative to the Nyquist rate. For a signal bandlimited to $W$ Hz, Nyquist sampling is $f_s=2W$, while oversampling by factor $\eta$ means $f_s=\eta\cdot 2W$; in particular, $\eta=2$ gives $f_s=4W$ (Koch et al., 2010). In wideband detection and related digital receiver models, the same idea is expressed as an oversampling ratio $M=f_s/B$, where $B$ is the signal bandwidth (Mariani et al., 2019). In correlation-coded OTDR, oversampling is parameterized by an integer ratio $M>1$0 through $M>1$1 and $M>1$2, where $M>1$3 is the code-bit width (Liao et al., 2017). In massive MIMO-OFDM with low-resolution ADCs, the oversampling ratio is written as $M>1$4, with FFT/IFFT size $M>1$5 and ADC sampling at $M>1$6 rather than the minimum Nyquist rate $M>1$7 (Ma et al., 2023).

A different but structurally analogous definition appears in imbalanced learning. There, oversampling refers to increasing the representation of the minority class beyond its empirical class frequency. The canonical form is SMOTE, which creates synthetic minority examples in feature space rather than simply replicating observations (Chawla et al., 2011). Deep SMOTE preserves that objective but replaces random interpolation with a learned regression model $M>1$8 (Mansourifar et al., 2020).

In multiscale finite elements, oversampling refers neither to time nor frequency. Instead, each coarse element $M>1$9 is associated with a larger connected patch $W$0, or more specifically with $W$1 coarse layers $W$2, and local corrector problems are solved on that enlarged domain (Henning et al., 2012). This suggests a broader unifying interpretation: oversampling is the intentional use of redundant local measurements or local support so that a coarse observation model can be compensated by additional structure.

2. Information-theoretic effects in communication channels

A central result for oversampling in quantized Gaussian channels is that doubling the sampling rate can recover part of the capacity loss caused by one-bit output quantization. For a continuous-time input bandlimited to $W$3 Hz with additive white Gaussian noise of two-sided PSD $W$4, followed by low-pass filtering to $W$5 Hz, hard limiting to $W$6, and sampling at rate $W$7, the relevant low-power metric is the capacity per unit-cost

$W$8

Without quantization, the classical result is $W$9. With one-bit uniform quantization and Nyquist sampling, the slope is

$f_s=2W$0

When the receiver oversamples at $f_s=2W$1, the derived lower bound becomes

$f_s=2W$2

raising the slope from $f_s=2W$3 to at least $f_s=2W$4, or about a $f_s=2W$5 gain (Koch et al., 2010). The mechanism is explicitly temporal: coarse one-bit quantization destroys amplitude information, but the bandlimited signal and filtered noise remain temporally correlated, so faster-than-Nyquist sign samples carry additional information in their joint law (Koch et al., 2010).

An analogous phenomenon appears in noncoherent Rayleigh block-fading channels at high SNR. Symbol matched filtering yields a discrete-time model with capacity pre-log

$f_s=2W$6

where $f_s=2W$7 is the number of symbols per fading block and $f_s=2W$8 is the rank of the covariance matrix of the discrete-time channel gains within the block. With oversampling by a factor two, the capacity pre-log is proved to satisfy

$f_s=2W$9

Hence symbol matched filtering is not capacity achieving for the underlying continuous-time channel (Dörpinghaus et al., 2014). The interpretation given is that multiplication by a time-varying fading process widens the receive bandwidth to $\eta$0, so symbol-rate sampling at rate $\eta$1 is not a sufficient statistic when $\eta$2 (Dörpinghaus et al., 2014).

These results place oversampling within a precise information-theoretic trade-off. In one-bit channels it compensates coarse amplitude quantization by finer time resolution (Koch et al., 2010). In noncoherent fading it compensates channel uncertainty by collecting extra temporal dimensions that allow the receiver to learn the fading waveform more efficiently (Dörpinghaus et al., 2014). A plausible implication is that oversampling is most consequential when the dominant loss mechanism is not thermal noise alone, but information discarded by a receiver architecture.

3. Low-resolution ADC systems and massive MIMO

Oversampling has been extensively analyzed as a mitigation strategy for one-bit ADC front ends. In large-scale MIMO uplink reception, conventional high-resolution ADC systems sample each antenna chain once per symbol interval, whereas one-bit ADCs introduce severe nonlinear distortion. Oversampling at $\eta$3 times the Nyquist rate partially compensates this loss by capturing more temporal information per symbol interval (Shao et al., 2021).

The 1-bit quantized receive model is commonly linearized via the Bussgang decomposition

$\eta$4

with $\eta$5 and distortion covariance $\eta$6 (Shao et al., 2021). In dynamic oversampling, the system first samples at rate $\eta$7, then applies a dimension-reduction matrix $\eta$8 so that only $\eta$9 samples per symbol interval are digitally processed:

$f_s=\eta\cdot 2W$0

The design criteria considered are sum-rate maximization and MSE minimization, and both lead to the same ratio-trace optimization,

$f_s=\eta\cdot 2W$1

which can be solved by generalized eigenvalue decomposition or by submatrix-level feature-selection heuristics (Shao et al., 2021). Simulations reported that dynamic oversampling outperforms uniform oversampling in computational cost, achievable sum rate, and symbol error rate performance (Shao et al., 2021).

For channel estimation in one-bit large-scale MIMO, oversampling is introduced as $f_s=\eta\cdot 2W$2 samples per symbol interval $f_s=\eta\cdot 2W$3, with receive vector dimension $f_s=\eta\cdot 2W$4. The oversampled matched-filtered model is

$f_s=\eta\cdot 2W$5

followed by element-wise one-bit quantization of the real and imaginary parts (Shao et al., 2019). Because oversampling induces colored noise after matched filtering, the paper derives Fisher information expressions for the white-noise case $f_s=\eta\cdot 2W$6 and a lower bound for colored noise when $f_s=\eta\cdot 2W$7 (Shao et al., 2019). The numerical result highlighted is that for $f_s=\eta\cdot 2W$8, $f_s=\eta\cdot 2W$9, and pilot length $\eta=2$0, $\eta=2$1 sits approximately $\eta=2$2 dB worse than $\eta=2$3 or $\eta=2$4 at the same NMSE, and the gains beyond $\eta=2$5 are diminishing (Shao et al., 2019).

In uplink massive MIMO-OFDM with low-resolution ADCs, oversampling ratio $\eta=2$6 enters directly into an achievable-rate approximation. With Bussgang coefficient $\eta=2$7 and $\eta=2$8, the simplified single-user MRC expression is

$\eta=2$9

where

$f_s=4W$0

The term $f_s=4W$1 models the residual quantization-distortion power after oversampling and averaging, so doubling $f_s=4W$2 halves that contribution (Ma et al., 2023). The reported design guidance is that with very low ADC resolution $f_s=4W$3–$f_s=4W$4, moderate OSR values $f_s=4W$5–$f_s=4W$6 yield substantial rate gains, whereas for $f_s=4W$7 bits or $f_s=4W$8 dB the benefit of $f_s=4W$9 is marginal (Ma et al., 2023).

The following comparison summarizes the principal communication-theoretic gains reported in the supplied literature.

Setting	Baseline	Oversampling result
One-bit Gaussian channel	$M=f_s/B$0	At $M=f_s/B$1, $M=f_s/B$2 (Koch et al., 2010)
Noncoherent Rayleigh block fading	$M=f_s/B$3	With $M=f_s/B$4 oversampling, $M=f_s/B$5 (Dörpinghaus et al., 2014)
1-bit MIMO channel estimation	$M=f_s/B$6	$M=f_s/B$7 or $M=f_s/B$8 gives about $M=f_s/B$9 dB SNR gain at same NMSE (Shao et al., 2019)
Massive MIMO-OFDM with low-res ADCs	$B$0	Quantization term scales as $B$1; gains strongest for $B$2–$B$3 (Ma et al., 2023)

Taken together, these results establish that oversampling can act as a substitute for ADC precision, pilot overhead, or matched-filter sufficiency, but the gains depend strongly on SNR regime, quantizer resolution, and whether the receiver can exploit the induced correlation structure.

4. Detection, sensing, and optical-fiber measurement

In oversampling-based wideband spectrum sensing, the sampling frequency exceeds the signal bandwidth, so that the discrete noise and any bandlimited signal become temporally correlated over roughly $B$4 consecutive samples (Mariani et al., 2019). This invalidates white-noise models of the form $B$5 and replaces them with full Toeplitz covariance matrices. The paper emphasizes two practical impairments in low-cost SDR front ends: noise-power uncertainty and non-flat noise PSD. Under these conditions, oversampling does not merely increase observation count; it changes the detection problem into one of exploiting covariance or spectral-shape departures from the noise-only hypothesis (Mariani et al., 2019).

Several detector classes are compared. The conventional energy detector suffers from noise uncertainty, while the ENP-ED statistic

$B$6

is invariant to unknown $B$7 (Mariani et al., 2019). For colored noise, the proposed PSD-matching detectors include a correlation-coefficient detector

$B$8

and a normalized KL detector defined on normalized periodograms and calibrated PSD templates (Mariani et al., 2019). The study reports that the best performance is provided by a noise-uncertainty immune energy detector and, for the colored-noise case, by tests that match the PSD of the receiver noise (Mariani et al., 2019). A moderate oversampling ratio $B$9 is described as typically sufficient, while excessive oversampling increases processing cost with diminishing returns (Mariani et al., 2019).

In correlation-coded OTDR, oversampling is used to improve coding gain without changing optical pulse width or injected energy, so spatial resolution remains unchanged (Liao et al., 2017). With oversampling factor $M>1$00, each code bit is replicated $M>1$01 times in the digital record, and the decoded OTDR response becomes an equivalent triangular pulse. For a length-$M>1$02 Golay code without oversampling, the correlation-decoded SNR improves over single-pulse OTDR by $M>1$03. Under ideal white noise and oversampling $M>1$04, the coding gain becomes

$M>1$05

Moreover, when the photodetector bandwidth $M>1$06 is finite and the sampling interval satisfies

$M>1$07

or equivalently

$M>1$08

the correlation of the filtered noise can produce an additional SNR boost beyond $M>1$09 (Liao et al., 2017). Experimental validation with 150 MHz and 300 MHz photodetectors showed the predicted optimal regions and reported that measured SNR gain follows the theory curves within $M>1$10–$M>1$11 dB (Liao et al., 2017).

These applications illustrate a recurring principle: oversampling is most useful when the analog front end already imposes memory, filtering, or correlation. The extra samples are then not redundant copies of identical information; rather, they expose dependence structures that can be exploited by an appropriately designed detector or decoder.

5. Data imbalance and synthetic oversampling in machine learning

In supervised learning on imbalanced datasets, oversampling refers to rebalancing class representation by increasing the minority class. The classical problem is that standard learners trained on skewed data tend to predict the majority class almost always, making predictive accuracy misleading and motivating ROC analysis, AUC, and ROC convex-hull analysis (Chawla et al., 2011).

SMOTE constructs synthetic minority examples in feature space instead of replicating minority observations. Given a minority-class feature vector $M>1$12, one of its $M>1$13 nearest minority neighbors $M>1$14, and $M>1$15, SMOTE generates

$M>1$16

The stated rationale is geometric: replication tends to carve out very small, highly specific regions and can lead to overfitting, whereas synthetic points between minority samples encourage larger and smoother minority-class decision regions (Chawla et al., 2011). In the experimental comparison with C4.5, Ripper, and Naive Bayes, the combination of SMOTE and majority under-sampling achieved better classifier performance in ROC space than under-sampling alone and better ROC convex-hull behavior than varying loss ratios or class priors (Chawla et al., 2011). The detailed summary reports that on 48 total experiments, SMOTE-based methods outperformed the alternatives in 44 cases (Chawla et al., 2011).

Representative AUC improvements with C4.5 were reported as follows (Chawla et al., 2011):

Dataset	Under-sampling AUC	SMOTE + Under-sampling AUC
Pima	0.626	0.655
Phoneme	0.658	0.728
Satimage	0.812	0.845
Forest cover	0.743	0.770
Oil	0.715	0.830
Mammography	0.603	0.668
E-state	0.739	0.796
Can	0.872	0.889

Deep SMOTE retains the interpolation principle but seeks to remove the instability induced by random sampling. Minority vectors $M>1$17 are paired, concatenated into $M>1$18, and mapped by a regression model

$M>1$19

Training minimizes the mean squared error to standard SMOTE targets $M>1$20 (Mansourifar et al., 2020). The principal claim is that once trained, $M>1$21 deterministically maps any pair of minority feature vectors to a synthetic point, yielding stable and reproducible oversampling (Mansourifar et al., 2020). The reported experimental finding is that Deep SMOTE can outperform traditional SMOTE in terms of precision, F1 score, and AUC in the majority of test cases (Mansourifar et al., 2020).

A common misconception is that oversampling in machine learning simply duplicates minority points. The supplied literature distinguishes replication from synthesis quite sharply. SMOTE is presented specifically as an alternative to naive over-sampling by replication, and Deep SMOTE is presented as an attempt to stabilize the synthetic generation step rather than to alter the fundamental minority-manifold interpolation idea [(Chawla et al., 2011); (Mansourifar et al., 2020)]. This suggests that, in data-level imbalance correction, the key issue is not only class frequency but also the geometry of the induced decision region.

6. Oversampling in deep networks, multiscale numerics, and digital resampling

The term also appears in deep neural networks in a sampling-theoretic sense. A feed-forward network is interpreted as a cascade of discrete-time linear filters followed by nonlinear activations, with each layer subject to the laws of sampling theory (Simpson, 2015). In this formulation, an oversampling factor $M>1$22 is defined at each layer, and $M>1$23 means the layer is over-sampled relative to the input bandwidth (Simpson, 2015). The stated mechanisms are that nonlinear activations generate harmonics and intermodulation terms, many of which alias back into the representable band when $M>1$24 is small, and that increasing $M>1$25 reduces these aliasing effects and yields more selective filters and more stable learning (Simpson, 2015). On decimated MNIST, over-sampled networks with factors $M>1$26 were compared with a baseline $M>1$27 network; the $M>1$28 model was reported to converge fastest and to the lowest error, about $M>1$29, while the layer-averaged crest factor was very strongly inversely correlated with error $M>1$30 (Simpson, 2015).

In the Multiscale Finite Element Method, oversampling refers to solving local corrector problems on patches larger than the target coarse element. The constrained-oversampling MsFEM introduces the coarse space

$M>1$31

the fine-scale kernel

$M>1$32

and patch-localized spaces

$M>1$33

For each coarse element $M>1$34 and coordinate direction $M>1$35, the local corrector $M>1$36 solves

$M>1$37

The central theorem states that if the patch depth satisfies $M>1$38, then the multiscale approximation satisfies an $M>1$39 error bound of the form

$M>1$40

with constants independent of oscillations of $M>1$41 or of $M>1$42 (Henning et al., 2012). The paper emphasizes that this provides the first rigorous proof of convergence for a MsFEM with oversampling and that the constrained construction eliminates resonance effects (Henning et al., 2012). Here oversampling means enlarged local computational context rather than increased temporal sample density, yet the purpose is similar: local redundancy is introduced so that a coarse discretization does not lose essential subscale information.

A further use of the term appears in digital resampling. Farrow filters are described as universal oversamplers for fractional delay correction and sampling-rate conversion (Bakholdin et al., 12 Mar 2025). The generic Farrow interpolator writes

$M>1$43

so oversampling by integer factor $M>1$44 is implemented by evaluating fractional phases $M>1$45 (Bakholdin et al., 12 Mar 2025). The paper contrasts cubic Lagrange Farrow filters, whose processing bandwidth is only about $M>1$46, with Hermite-spline Farrow structures of order $M>1$47, $M>1$48, and $M>1$49, which increase the processing bandwidth up to $M>1$50 of sampling frequency and improve continuity and sidelobe performance (Bakholdin et al., 12 Mar 2025). This application concerns the construction of oversamplers themselves rather than the use of oversampling within a larger inference task.

7. Trade-offs, limitations, and recurring themes

Across the supplied literature, oversampling is consistently associated with a trade-off between additional measurements and additional cost. In one-bit MIMO receivers, oversampling increases ADC throughput, memory use, and baseband processing burden, motivating dynamic oversampling schemes that keep the digital processing rate at a smaller $M>1$51 even when the analog sampling rate is higher (Shao et al., 2021). In channel estimation, the complexity of the low-resolution-aware LS estimator grows as $M>1$52, and receiver front ends must support higher RF/IF and digital processing bandwidth as $M>1$53 increases (Shao et al., 2019). In massive MIMO-OFDM, ADC sampling power grows linearly in $M>1$54, while the sum-rate gain grows only logarithmically through the effective SINDR (Ma et al., 2023). In wideband sensing, moderate oversampling ratios are favored because larger $M>1$55 yields diminishing returns relative to processing cost (Mariani et al., 2019). In OTDR, the method is attractive precisely because only the ADC sampling rate is adjusted, without changes to the optical transmitter (Liao et al., 2017).

The limitations are domain specific but conceptually similar. The one-bit Gaussian-channel result is derived only in the very-low-power regime $M>1$56 and only for binary inputs and binary outputs; gains beyond $M>1$57 and for nonbinary quantization remain open (Koch et al., 2010). In noncoherent fading, the result is a lower bound on pre-log rather than a full finite-SNR capacity characterization (Dörpinghaus et al., 2014). In SMOTE, oversampling may increase class overlap and false positives when minority features have large variance or overlap heavily with majority space (Chawla et al., 2011). Deep SMOTE may degrade on very low-dimensional data such as Haberman (Mansourifar et al., 2020). In deep neural networks, the robustness claim is presented as a prediction from the sampling-theoretic argument rather than as a directly measured adversarial robustness result (Simpson, 2015). In MsFEM, accuracy depends on patch depth and the constrained local formulation, not merely on enlarging patches indiscriminately (Henning et al., 2012).

A recurring misconception is that oversampling simply means “more samples are always better.” The literature does not support that simplistic view. In communications, gains arise because oversampling exploits temporal correlation left by bandlimitation, filtering, or channel variation, not because independent information is created ex nihilo [(Koch et al., 2010); (Dörpinghaus et al., 2014)]. In one-bit ADC architectures, the best designs often combine oversampling with dimension reduction or structured sample selection, indicating that the value lies in which extra samples are retained and how they are processed (Shao et al., 2021). In imbalanced learning, synthetic oversampling is differentiated from naive replication because geometry, not only cardinality, determines classifier behavior (Chawla et al., 2011). In multiscale numerics, oversampling patches are useful only when coupled to appropriate constraints that isolate fine-scale correctors from coarse basis functions (Henning et al., 2012).

Viewed across fields, oversampling is best understood as a method for compensating a bottleneck imposed elsewhere in the system. That bottleneck may be one-bit quantization, symbol-rate sufficiency assumptions, skewed empirical class priors, aliasing induced by nonlinear activations, or coarse finite-element localization. The consistent research finding is that redundancy in time, feature space, or computational support can partially restore lost information, but only when the induced dependencies are modeled rather than ignored [(Koch et al., 2010); (Dörpinghaus et al., 2014); (Chawla et al., 2011); (Henning et al., 2012)].