Signal-to-noise ratio is the ratio of the power of a desired signal to the power of accompanying noise, evaluated under specified conditions. It is a fundamental measure in telecommunications, signal processing, and experimental measurement because it helps determine whether information can be detected, estimated, transmitted, or reproduced reliably. The ratio is dimensionless and is commonly expressed on the logarithmic decibel scale.
A meaningful signal-to-noise ratio requires a definition of the desired signal, the unwanted fluctuations, the measurement bandwidth, and the point in a system at which the comparison is made. A radio receiver, an image detector, and a chemical spectrometer may therefore report superficially similar numbers that describe different physical quantities. A high ratio generally makes information recovery easier, but it does not by itself establish accuracy, fidelity, or freedom from systematic error.
The concept connects physical noise mechanisms with statistical inference and engineering design. Thermal agitation, the discrete arrival of photons and electric charges, electronic amplification, and numerical quantization all create different limitations. Their consequences depend not only on total noise power but also on its distribution in frequency, time, and space, and on the processing used to extract the signal.
Definition, conventions, and mathematical interpretation
The conventional power definition is
where is desired-signal power and is noise power. Its logarithmic form is
Thus, 0 dB means equal signal and noise powers, 10 dB means a power ratio of 10, and 30 dB means a ratio of 1,000. Negative values are physically meaningful: at −10 dB, noise power is ten times signal power. Such a signal can still be recoverable when its structure is known or sufficient observations are available.
For voltages measured across the same resistance, power is proportional to mean-square voltage, giving
The factor of 20 applies because a squared amplitude is being converted into a power ratio. It is not a separate definition of the decibel. If impedances differ, voltage ratios alone do not determine power ratios. Similar care is necessary with acoustic pressure, optical field amplitude, and digital sample values.
A common model represents a measurement as
where is the desired waveform and is additive noise. For stationary processes, mean-square signal and noise values can define their respective powers. If both are zero-mean and uncorrelated, their powers add:
Correlation introduces a cross term, so this identity is not universal. Moreover, mean square and variance are equal only for a zero-mean quantity. A constant voltage has zero variance but nonzero electrical power, illustrating why a signal’s physical meaning matters.
In imaging, spectroscopy, and other measurement fields, “SNR” frequently means a mean signal amplitude divided by a noise standard deviation:
This amplitude-based quantity is not numerically identical to a power ratio. Where the corresponding powers are proportional to and , the power ratio is . Published values therefore require attention to the stated convention.
Signal and noise are also task-dependent categories. A periodic component may be the desired signal in a vibration measurement but an unwanted artifact in another experiment. The distinction is operational rather than an intrinsic property of a waveform. Likewise, a systematic calibration offset may produce little random noise while making the measurement inaccurate; SNR alone cannot detect that failure.
Noise mechanisms, spectra, and measurement bandwidth
Noise power depends on the physical mechanism that generates fluctuations and on the frequency range admitted by the measurement. Thermal noise, also called Johnson–Nyquist noise, arises from thermal motion of charge carriers. John B. Johnson reported experimental measurements in 1928, and Harry Nyquist developed the associated theoretical explanation that year. In the classical regime, the available noise power from a matched source is
where is the Boltzmann constant, is absolute temperature, and is bandwidth. This approximation applies when photon energy is small relative to thermal energy, .
For an ideal resistor of resistance , the open-circuit mean-square noise voltage over bandwidth is . At 290 K, available thermal noise has a spectral density of approximately −174 dBm/Hz. Integrating that density over 1 MHz gives about −114 dBm before receiver-added noise. Here dBm expresses power relative to 1 milliwatt; unlike SNR in decibels, it represents an absolute power level.
Shot noise results from the discrete transfer or detection of charge and light. For an ideal uncorrelated electric current , a common one-sided current-noise spectral density is , where is elementary charge. In optical detection, independent photon arrivals often follow Poisson statistics. If the mean detected count is , the standard deviation is . The amplitude-based SNR of a signal-only count is then , not . This dependence makes improvements increasingly expensive in photon-limited measurements.
Other noise mechanisms have different spectra. Flicker noise has a power spectral density approximately proportional to , with often near one over a limited range. It is important in low-frequency electronics and precision measurements. Impulsive noise contains occasional large disturbances and may be poorly described by its average power. Environmental pickup, vibration, and power-line interference can introduce narrow spectral peaks rather than a smooth background.
The general connection between spectrum and total noise is expressed through power spectral density. For a filter with frequency response , using a consistent two-sided convention,
Thus, a filter’s shape matters, not merely its nominal cutoff frequencies. The equivalent noise bandwidth is the bandwidth of an ideal rectangular filter that would transmit the same white-noise power at the specified reference gain. A narrower bandwidth reduces white-noise power, but it can also remove signal components, distort pulses, or lengthen the response time. Bandwidth reduction improves SNR only when those costs are compatible with the measurement.
Receiver noise, cascaded systems, and sensitivity
A receiver adds noise as well as amplifying the signal and the noise already present at its input. Its degradation of SNR is described by the noise factor,
Noise figure is the logarithmic expression . An ideal noiseless amplifier has , or a noise figure of 0 dB. A real amplifier can increase signal amplitude substantially while reducing SNR, because its internal devices contribute additional fluctuations. Gain and SNR improvement are therefore distinct properties.
Conventional noise-factor specifications use a reference source temperature, ordinarily 290 K. An alternative description is equivalent input noise temperature:
where is the reference temperature. Noise temperature is particularly useful in radio astronomy, satellite links, and low-noise receiving systems, where the antenna may view a background quite different from a room-temperature termination. Antenna noise, losses ahead of the amplifier, and receiver-added noise together determine system sensitivity.
Harald T. Friis published the standard cascade relation in 1944. For stages with linear power gains and noise factors ,
The equation explains why the first stages are especially important: sufficient low-noise gain at the input reduces the input-referred contribution of subsequent stages. The gains and noise factors in this expression are linear quantities, not decibel values.
Loss before the first amplifier can be particularly damaging. A passive attenuator at the reference temperature has a noise factor equal to its linear power loss under the conventional matched conditions. A 3 dB loss ahead of a receiver therefore contributes approximately 3 dB of noise figure by itself. At other physical temperatures, the attenuator’s thermal emission changes, so its noise contribution must be calculated accordingly. Cryogenic receivers exploit reduced thermal noise, but cooling does not remove sky background, quantum fluctuations, or every device-noise mechanism.
Receiver sensitivity is often estimated by combining thermal noise, bandwidth, noise figure, and the SNR required by a particular task. Near 290 K, a simplified expression is
For a 1 MHz bandwidth, 5 dB noise figure, and 10 dB required SNR, this gives approximately −99 dBm. The result is conditional rather than a universal receiver rating. Strong adjacent-channel signals, oscillator noise, overload, and nonlinear distortion can impair reception even when the desired signal exceeds this thermal-noise-based threshold.
Measurement procedures and reporting conventions
SNR measurement begins by specifying a reference plane: the physical or numerical location at which signal and noise are compared. Input SNR, amplifier-output SNR, demodulator SNR, and reconstructed-image SNR generally differ. The operating temperature, bandwidth, termination, averaging interval, signal level, and processing settings may all affect the result. An isolated decibel value without these conditions is often insufficient for comparison.
When signal and noise can be observed separately, their powers can be measured directly. In a signal-on/signal-off procedure, the signal-on measurement contains both contributions:
If the noise is unchanged when the signal is removed, then
Subtracting or dividing decibel readings directly does not perform this power subtraction; the readings must first be converted to linear units. The method also fails if switching the source changes the system’s noise, as can happen with signal-dependent shot noise, automatic gain control, or source heating.
For sinusoidal electrical tests, a spectrum or discrete Fourier transform can separate a tone from surrounding noise. The analysis must identify which bins belong to the desired tone, which contain harmonics or spurious signals, and which contribute to the noise estimate. If the sampled record does not contain an integer number of tone periods, spectral leakage spreads tone energy into adjacent bins. Window functions reduce leakage but alter the effective noise bandwidth and amplitude calibration.
A displayed spectral noise floor is not itself the total noise power. For approximately white noise, reducing an analyzer’s resolution bandwidth lowers the power in each frequency bin. The integrated noise over a fixed overall bandwidth does not fall merely because more, narrower bins are displayed. Detector mode, logarithmic averaging, and window normalization can also change a displayed floor, so comparisons require equivalent settings.
Finite data introduce statistical uncertainty. Noise power estimates fluctuate, especially when based on short records or few independent averages. Averaging spectra reduces the variability of the estimate but does not necessarily reduce the physical noise in the measured system. Correlated samples provide fewer independent observations than their raw count suggests, connecting SNR measurement directly with statistics and experimental design.
Several neighboring metrics answer different questions. Signal-to-noise-and-distortion ratio includes distortion as well as noise. Dynamic range describes an available range between limiting levels, often a maximum usable signal and a noise or distortion floor. Peak signal-to-noise ratio compares reconstruction error with a specified peak signal value. These measures can be useful, but substituting one for another can conceal different test conditions or performance limitations.
Detection, estimation, and processing gain
SNR affects detection, but detection performance also depends on what is known about the signal and noise. A detector normally balances the probability of detecting a real signal against the probability of declaring a signal when none is present. Two systems with the same input SNR can perform differently if one knows the signal’s arrival time and waveform while the other must search many possible frequencies, delays, or spatial positions. A larger search introduces more opportunities for noise to imitate a signal.
For a known deterministic waveform in additive white noise, the matched filter maximizes output SNR at a selected sampling instant among linear filters. Its impulse response is proportional to a time-reversed, complex-conjugated version of the waveform. The filter adds contributions that follow the expected signal pattern while allowing unrelated fluctuations to cancel statistically. This is a central technique in radar, communications, and pulse detection.
Using a real-signal convention with two-sided white-noise spectral density , the maximum matched-filter output SNR for a signal of energy is
Different real or complex baseband conventions can move factors of two between definitions, making consistent notation essential. For colored noise, the optimum linear detector accounts for the noise spectrum as well as the signal waveform; frequencies with greater noise receive less weight.
Repeated observations provide another form of processing gain. Suppose
where the noise samples are independent, zero-mean, and have variance . The average of observations has noise variance . Its power SNR improves by , or dB, while its amplitude-to-standard-deviation ratio improves by . Averaging 100 independent observations therefore provides 20 dB of power-SNR improvement. Correlated drift or a persistent offset does not decrease according to this rule.
In radar, a long coded or frequency-swept pulse can carry more energy than a short pulse at the same peak power. Matched filtering compresses the received waveform in time, allowing substantial energy collection while retaining fine range discrimination. The result depends on pulse energy, bandwidth, waveform sidelobes, and target behavior; pulse compression does not create information or energy.
Estimation problems require still more than total SNR. The uncertainty of a time-delay estimate depends strongly on signal bandwidth, while frequency-estimation precision depends on observation duration and waveform structure. Fisher information and the Cramér–Rao bound formalize these relationships under specified statistical models. A narrowband tone and a broadband pulse can have identical power SNRs yet support very different timing accuracy. SNR is therefore an important input to inference, not a complete description of inferential performance.
Communication capacity, modulation, and interference
In communication systems, SNR links the physical quality of a channel with the rate at which information can be conveyed. Claude Shannon’s 1948 work established that an ideal band-limited additive white Gaussian noise channel has capacity
where is capacity in bits per second, is bandwidth in hertz, and SNR is a linear received-power ratio. This Shannon–Hartley relation describes an asymptotic limit under its channel and power assumptions. It is not the guaranteed throughput of a particular modem.
The equation shows diminishing returns from increasing signal power. At a fixed bandwidth of 1 MHz and SNR of 10, corresponding to 10 dB, capacity is approximately 3.46 Mbit/s. Raising SNR to 100, or 20 dB, increases capacity to about 6.66 Mbit/s rather than multiplying it by ten. Practical systems achieve lower net rates because of coding limitations, synchronization, pilots, framing, retransmissions, and other overheads.
Digital communication frequently uses energy per information bit relative to noise spectral density:
where is information bit rate and noise power is under the stated convention. This separates energy efficiency from bandwidth use. If spectral efficiency is , the ideal capacity boundary gives
As approaches zero, the limiting value approaches , approximately −1.59 dB. This limit requires increasingly low spectral efficiency and idealized coding assumptions; it is not a general operating threshold for radios.
The required SNR also depends on modulation and error-control coding. Higher-order constellations transmit more bits per symbol but place candidate symbols closer together for a given average power, generally requiring higher SNR for the same uncoded error rate. For coherent binary phase-shift keying in additive white Gaussian noise,
where is the upper-tail probability of a standard normal distribution. Error-correcting codes change the relationship between received SNR and recovered-bit error probability by adding structured redundancy.
Unwanted transmissions require a further distinction. Signal-to-interference-plus-noise ratio, or SINR, includes interference power in the denominator. Structured interference may not behave like Gaussian noise, even when the total power is identical. Wireless links also experience fading, so instantaneous SNR varies with time and location; average SNR alone can conceal a substantial probability of outage. Adaptive modulation, coding, and spatial processing respond to these variations rather than treating the channel as a single fixed ratio.
Digital conversion, audio, and numerical representation
An analog-to-digital converter introduces finite-resolution error when it maps a continuous input onto discrete output levels. For an ideal uniform quantizer with step size , a widely used model treats the error as uniformly distributed between and . Its mean-square value is then
The model is useful when the signal traverses many levels and quantization error behaves approximately like uncorrelated noise. It is not exact for every input: a constant value or a small periodic waveform may produce deterministic error patterns rather than a white-noise spectrum.
For an ideal -bit converter driven by a full-scale sine wave, the resulting signal-to-quantization-noise ratio is approximately
The familiar improvement of roughly 6 dB per bit follows from halving the quantization step with each additional bit. The expression assumes an ideal converter, the specified full-scale sinusoidal input, and integration over the relevant full Nyquist band. It does not state that all -bit recordings or measurements possess that SNR. Reducing the signal amplitude lowers its ratio to an unchanged quantization-noise floor.
Real converters also contain thermal noise, distortion, reference noise, and clock-related errors. Effective number of bits is often inferred from signal-to-noise-and-distortion ratio using
Because this measure includes distortion, it differs from a bit count inferred from random noise alone. Input frequency matters as well. In a simplified model, root-mean-square sampling-time jitter limits the SNR of a sinusoid at frequency approximately to
The same clock jitter is therefore more damaging at higher input frequencies.
Oversampling spreads approximately white quantization noise across a wider frequency range relative to the useful signal band. Subsequent digital low-pass filtering can remove out-of-band noise. Under the ideal white-noise model, doubling the sampling rate while retaining a fixed signal bandwidth improves in-band quantization SNR by about 3 dB. Noise shaping redistributes quantization noise toward frequencies outside the useful band, permitting larger in-band improvements under suitable conditions.
In sound recording, SNR is one aspect of fidelity rather than a complete perceptual score. Hum, clicks, harmonic distortion, clipping, and frequency-response errors can be objectionable even when integrated random-noise power is low. Frequency weighting changes the reported ratio by emphasizing some parts of the audible spectrum more than others. Dither deliberately adds a small amount of noise to reduce signal-correlated quantization artifacts; it can improve low-level linearity while increasing measured noise power.
Optical detection, imaging, and scientific instruments
Many optical and imaging instruments measure discrete detection events rather than a continuous electrical waveform. For independent photoelectron counts, a useful amplitude-SNR model is
where is the mean signal count, is background count, is dark-current count, and is readout-noise variance in squared-electron units. The expression assumes independent contributions and does not include every calibration or reconstruction error. Subtracting the mean background removes its offset but not the shot noise generated during detection.
When signal shot noise dominates, SNR scales as . When readout noise dominates a single exposure, SNR is approximately . These regimes explain why exposure time, optical throughput, and repeated readouts have different consequences. Combining many short exposures incurs read noise repeatedly, whereas a single longer exposure can avoid that penalty but may encounter saturation, motion, dark current, or transient contamination.
A photodiode converts detected light into charge or current; a photomultiplier tube adds internal multiplication to make weak events easier to measure. Multiplication can reduce the relative importance of later electronic read noise, but stochastic gain can introduce additional noise. Greater output amplitude is therefore not equivalent to a proportional improvement in input-referred SNR. Detector quantum efficiency—the fraction of incident photons producing useful detections—is often as important as subsequent gain.
In imaging, SNR alone does not specify whether an object can be distinguished from its surroundings. Contrast-to-noise ratio compares an intensity difference with a defined noise scale. Spatial resolution also matters: averaging neighboring pixels lowers random noise but can erase a small structure. A smooth image may have a higher numerical SNR while preserving less fine detail. Nonlinear reconstruction further complicates the comparison because it can make noise spatially varying and correlated.
In magnetic resonance imaging, receive-coil sensitivity, voxel volume, acquisition bandwidth, pulse sequence, and measurement time all influence SNR. Combining receiver channels changes the noise distribution, and magnitude reconstruction introduces a positive bias at low signal levels. Consequently, noise estimated from an apparently empty image region does not always represent noise in tissue. In spectroscopy, a reported peak-height-to-baseline-noise ratio similarly depends on spectral resolution, line width, baseline treatment, and integration time.
The physical limits extend to quantum mechanics. Optical measurement noise can include both photon-counting fluctuations and disturbance associated with measurement. Squeezed optical states can reduce fluctuations in a chosen field quadrature while increasing fluctuations in the conjugate quadrature. Such techniques do not eliminate quantum uncertainty; they redistribute it in a way suited to a particular measurement.
Improving SNR and recognizing its limits
SNR can be improved by increasing the useful signal, reducing unwanted fluctuations, or exploiting structure that distinguishes the two. These strategies operate at different stages. A larger collecting aperture, greater optical efficiency, or better coupling can increase detected signal energy. Shielding, differential measurement, thermal stabilization, and low-noise electronics can reduce specific disturbances. Their effectiveness depends on the dominant mechanism: electromagnetic shielding does not remove photon shot noise, and cooling a detector does not eliminate background light.
Frequency-selective methods are particularly useful when the signal can be moved into a quieter region. A lock-in amplifier compares an input with a reference at a known modulation frequency, then applies low-pass filtering to retain the coherent component. This can reject broadband noise and unrelated frequencies while avoiding some low-frequency drift. The improvement depends on integration time and equivalent noise bandwidth. Interference synchronous with the reference can survive the same processing.
When multiple sensors observe related signals, spatial processing can improve separation. An antenna array can align phases from a desired direction so that the signal adds coherently while some receiver noise adds incoherently. Under ideal assumptions of identical signal reception and independent equal-power receiver noises, combining elements can improve power SNR by . Real gains depend on array geometry, mutual coupling, calibration, and noise correlation. Correlated interference requires more than simple averaging.
Filtering and denoising always embody assumptions. A linear time-invariant filter cannot distinguish signal and noise solely by their labels; it acts on both according to frequency. A method that suppresses frequencies containing noise also suppresses useful signal occupying those frequencies. Statistical estimators and machine learning systems can exploit richer patterns, but their output depends on model assumptions or training data. A visually cleaner reconstruction may remove genuine features or introduce plausible ones that were not established by the observations.
Improvement is also constrained by dynamic range and nonlinear behavior. Raising transmitter power or detector exposure can improve a noise-limited measurement until amplification saturates, a detector reaches its full-well capacity, or another system limit becomes dominant. Strong out-of-band inputs can drive a receiver into nonlinear operation even though they would have been rejected by an ideal linear filter. Under those conditions, an SNR calculated from small-signal noise measurements no longer predicts performance reliably.
At the fundamental level, deterministic processing of a complete observation cannot create new information about an unknown signal. Processing can nevertheless produce a much better task-specific statistic by discarding irrelevant components or combining redundant observations. The appropriate test of an SNR improvement is therefore tied to the task: lower estimation uncertainty, better detection at a fixed false-alarm rate, fewer communication errors, or more reliable preservation of image detail. Different tasks can rank the same processed outputs differently, even when a single reported SNR increases.