From Oxide to Oversampling: The Physics of Recorded Sound
Abstract
‘Analogue warmth’ and ‘digital coldness’ are not aesthetic preferences alone — they involve different physics. Magnetic tape saturates and compresses in level- and machine-dependent ways. Delta-sigma modulators push much quantisation noise out of band. Both effects are calculable.
Contents
There is an argument that has been running in recording studios since roughly 1982, when the first commercially mastered compact discs appeared. On one side: analogue tape has warmth, depth, something the ear likes. On the other: digital audio is more accurate, lower noise, the measurements say so. The argument produces more heat than light, because most participants treat it as an aesthetic question — a matter of feeling, taste, preference. It is not only that. The difference between tape and digital audio includes physics specific enough to calculate.
The physics here turns out to be some of my favourite kind: it sits at the intersection of condensed matter, signal processing, and Fourier analysis. Part I is the ferromagnetic physics of magnetic tape and the harmonic structure of saturation distortion. Part II is the delta-sigma modulator and the engineering trick that achieves wide dynamic range from low-resolution quantisation. Neither side of the debate is as simple as its partisans claim.
Literature cutoff: 11 July 2026.
Part I: The Physics of Magnetic Tape
Ferromagnetic Recording
Magnetic recording tape is a thin polymer substrate coated with a layer of ferromagnetic particles suspended in a binder. For most of the twentieth century those particles were iron oxide — specifically $\gamma\text{-Fe}_2\text{O}_3$, gamma-phase ferric oxide — though chromium dioxide ($\text{CrO}_2$) and later metal-particle formulations with pure iron or iron-cobalt alloys were developed for higher coercivity and better high-frequency response. What all of these materials share is the key property of ferromagnetism: each particle is a small permanent magnet, a magnetic domain with a net magnetic moment that can be oriented by an external field and that will retain that orientation when the field is removed.
The recording process exploits this directly. The recording head is a toroidal electromagnet with a narrow gap. When audio-frequency current flows through the head’s coil, the field at the gap follows the current, and as the tape moves past at a fixed speed, successive particles along the tape length are aligned according to the instantaneous field at the moment they pass the gap. The result is a spatial encoding of the time-domain audio signal along the tape. On playback, the inverse process occurs: the moving pattern of magnetised particles generates a time-varying flux in the playback head’s core, which induces a voltage in the coil by Faraday’s law. That voltage is “proportional to the rate of change of the flux” [[6]], so it rises 6 dB per octave, and the playback amplifier’s equalisation undoes the slope to reproduce the original current waveform.
So far this description is entirely linear. The head current maps to a field, the field maps to a magnetisation, the magnetisation maps back to a voltage. If all three relationships were linear, tape would be a near-perfect recording medium — limited only by particle noise and head gap frequency response. The nonlinearity comes from the second relationship in that chain, and it comes from the fundamental physics of how ferromagnetic materials respond to an applied field.
The B-H Curve and Hysteresis
The relationship between the applied magnetic field intensity $H$ (from the recording head, measured in A/m) and the resulting magnetic flux density $B$ in the tape (measured in tesla) is not linear. It follows a curve — actually a family of nested curves — known as the hysteresis loop, and its shape determines almost everything interesting about tape recording [[3]].
Starting from a demagnetised state and increasing $H$ from zero, the initial slope $dB/dH$ — the magnetic permeability $\mu$ — is relatively low. The domains in the material are oriented randomly and require a threshold of energy to begin reorienting. As $H$ increases further, the permeability rises, and there is a region of steep, approximately linear increase in $B$. Then, as $H$ continues to increase, the material saturates: progressively fewer unaligned domains remain, the slope falls, and eventually $dB/dH \to 0$ as all domains are aligned. The $B$-$H$ curve is S-shaped, and the saturation is irreversible in a specific sense: if you now reduce $H$ back toward zero, $B$ does not retrace the original path. It remains at a higher value — the remanence $B_r$ — and you must apply a reverse field of magnitude $H_c$, the coercivity, to bring $B$ back to zero. The loop formed by this cycle of magnetisation and demagnetisation is the hysteresis loop, and its area is proportional to the energy dissipated as heat per cycle.
The crucial feature for audio recording is what happens near the origin. A small audio signal, sitting near $H = 0$, does not experience a nicely linear region of the $B$-$H$ curve. The initial permeability is low, and there is an inflection point near zero: the slope increases as you move away from zero before the saturation region brings it back down again. The transfer function is nonlinear. An ideally symmetric transfer function is odd under $H \to -H$, so its Taylor expansion contains odd powers and generates odd harmonics; even-order components require asymmetry, such as an offset or imperfect bias. Something needed to be done about this, and AC bias was the decisive practical solution.
The Bias Signal
The solution is called AC bias. It is usually credited to Walter Weber and Hans Joachim von Braunmühl at the German Reichs-Rundfunk-Gesellschaft, where in about April 1940 a test circuit “began to oscillate” and “a dramatic decrease in background noise was observed” [[7]]. There were earlier related patents, among them Carlson and Carpenter’s in the United States (1927) [[7]].
The idea is simple once stated: add a high-frequency signal, well above the audio band, to the recording current before it drives the head. A BASF rule of thumb puts it at least five times above the highest frequency to be recorded [[8]]. In front of the gap, its field exceeds the coercivity of the tape.
The effect on the recording process is to linearise the transfer function, though not by simple averaging. A particle approaching the gap is driven round a hysteresis loop by the growing bias field, shifted slightly by the much slower audio field. As it leaves, the bias amplitude decays over many cycles until the shrinking loop can no longer reverse the particle, and its magnetisation freezes. In the simplified model of a 1955 lecture by Philips’s W. K. Westmijze, the decisive field “is the field that exists at the moment that the minor hysteresis loop just closes”, and the remanence that results is “strictly proportional” to the audio field [[6]]. Too little bias and the linearisation is incomplete; too much and the high-frequency output falls through “particle self-erasure and spacing loss” [[8]]. Getting the bias right is part of the alignment procedure for every analogue tape machine and part of why different tape formulations require different machine settings.
The result, for moderate recording levels, is a remarkably clean and linear recording medium. The nonlinear character of the B-H curve is effectively tamed by the bias trick, and the remaining imperfections are mostly of other kinds: azimuth errors, print-through, head bump, self-demagnetisation at short wavelengths. A current studio tape, RTM’s LPR35, specifies 0.44 per cent third harmonic distortion at its reference level [[9]]. For practical purposes, a well-aligned analogue tape machine at moderate recording levels is a linear system to within a fraction of a per cent.
Harmonic Generation at High Levels
At high recording levels — when the audio signal is large enough to push the operating point into the saturation region even after the bias has done its linearising work — the picture changes. The transfer function from input current to output magnetisation becomes genuinely nonlinear, and the harmonic content of the distortion becomes the central question.
The standard framework is a Taylor expansion of the transfer function around the operating point:
$$y(t) = a_1 x(t) + a_2 x^2(t) + a_3 x^3(t) + a_4 x^4(t) + \cdots$$where $x(t)$ is the input signal (the audio current), $y(t)$ is the output (the magnetisation recorded on tape), and the coefficients $a_n$ are determined by the shape of the B-H curve near saturation. For a pure tone $x(t) = A \sin(\omega t)$, the higher-order terms generate harmonics in a calculable way.
The second-order term gives:
$$a_2 x^2(t) = a_2 A^2 \sin^2(\omega t) = \frac{a_2 A^2}{2}\bigl(1 - \cos 2\omega t\bigr)$$This is a DC offset plus a component at $2\omega$ — the second harmonic, one octave above the fundamental.
The third-order term gives:
$$a_3 x^3(t) = a_3 A^3 \sin^3(\omega t) = a_3 A^3 \left(\frac{3}{4}\sin\omega t - \frac{1}{4}\sin 3\omega t\right)$$The $\frac{3}{4}$ piece adds to (or subtracts from) the fundamental depending on the sign of $a_3$; the $-\frac{1}{4}$ piece is a third harmonic at $3\omega$, one octave and a fifth above the fundamental.
Carrying through to fourth order:
$$a_4 x^4(t) = \frac{a_4 A^4}{8}\bigl(3 - 4\cos 2\omega t + \cos 4\omega t\bigr)$$which contributes additional DC, a component at $2\omega$, and a fourth harmonic at $4\omega$.
Collecting all terms of the fourth-order polynomial, the output is:
$$ \begin{aligned} y(t) = {}& \frac{a_2 A^2}{2} + \frac{3a_4 A^4}{8} + \left(a_1 + \frac{3a_3 A^2}{4}\right)A\sin\omega t \\ &- \left(\frac{a_2 A^2}{2} + \frac{a_4 A^4}{2}\right)\cos 2\omega t - \frac{a_3 A^3}{4}\sin 3\omega t + \frac{a_4 A^4}{8}\cos 4\omega t \end{aligned} $$The important observation is about which harmonics dominate and what they sound like. The B-H curve of a ferromagnetic material near saturation is approximately symmetric: the saturation behaviour for positive $H$ mirrors that for negative $H$. A symmetric nonlinearity has $a_2 = a_4 = 0$, so it generates odd harmonics. In well-biased tape recorders, third harmonic is commonly the principal measured harmonic-distortion term near saturation; the LPR35 data sheet specifies distortion only as third harmonic, and sets its reference bias for “the minimum third harmonic distortion ratio” [[9]]. Even harmonics can appear when the record chain or bias is asymmetric. The exact spectrum depends on tape formulation, bias setting, head, level, and frequency. There is no universal “tape equals even harmonics” signature.
A common choice of smooth curve with this symmetry and a ceiling is the hyperbolic tangent,
$$y = \tanh x = x - \frac{x^3}{3} + \frac{2x^5}{15} - \cdots, \tag{1}$$with the input $x$ in units of the knee: small-signal gain 1, ceiling 1. It is a stand-in, not a model of tape, since it ignores hysteresis, bias and frequency, but it shows how a soft knee distributes its distortion. With $a_1 = 1$ and $a_3 = -1/3$, the display above predicts a third harmonic of $A^2/12$ relative to the fundamental for small $A$, −61.6 dB at $A = 0.1$, which a Fourier transform of the full curve reproduces. The series converges only for $|x| < \pi/2$, so Fig. 1 uses the full curve, not the series. At $A = 0.5$ the third harmonic sits at −34 dB; at $A = 2$, well into the knee, it has risen to about −15.5 dB, and each further odd harmonic is about 12.5 dB below the one before.
The perceptual consequence is less tidy than the usual “warmth” story. The second and fourth harmonics are octaves of the fundamental, and small amounts can be perceived as a change in timbre rather than roughness. Odd harmonics are not automatically harsh: the third is an octave plus a fifth, while high levels of any nonlinear distortion can sound abrasive. Audibility depends on spectrum, level, programme material, listening level, and the rest of the recording chain.
There is one more effect worth naming: the saturation is a soft knee. The B-H curve does not have a sharp corner at saturation — it curves gradually from the linear region into the flat-topped saturation region. This means that transient signals — percussive attacks, consonant onsets — that briefly exceed the nominal recording level are not hard-clipped but gently compressed. Their peaks are rounded by the shape of the B-H curve. Engineers and producers who record through tape often describe this as the machine “breathing” or as a pleasing “gluing” of transients. The physics is simple: the soft-knee transfer function applies more gain reduction to instantaneous peaks than to the sustained body of the signal. It acts as an instantaneous limiter that lowers the crest factor of anything approaching saturation. Unlike a compressor it has no attack or release time: it reshapes each cycle of the waveform, so the gain reduction arrives together with the harmonics of Fig. 1.
Part II: The Physics of Delta-Sigma Conversion
Nyquist-Rate ADC and Its Limits
The straightforward approach to analogue-to-digital audio conversion samples the signal at a rate just above twice the highest audio frequency — the Nyquist rate — using a quantiser with enough bits to achieve the desired dynamic range. For CD-quality audio, the sampling rate is 44.1 kHz (slightly above $2 \times 20{,}000$ Hz) and the word length is 16 bits. The dynamic range of a $b$-bit PCM system is, to a good approximation:
$$\text{SNR} \approx 6.02b + 1.76 \text{ dB}$$so 16 bits gives approximately $6.02 \times 16 + 1.76 \approx 98$ dB for a full-scale sine, well above the approximately 70 dB achievable with the noise floor of typical studio tape at 15 ips [[4]]. Westmijze’s 1955 estimate and the LPR35 data sheet agree on about 70 dB between maximum signal and noise [[6]], [[9]].
The engineering problem with a straightforward Nyquist-rate ADC is the anti-aliasing filter. Energy between $f_s/2 = 22.05$ kHz and $f_s$ that survives the filter reappears at $f_s - f$, inaudible in origin but very much audible in its alias. As the post on the 44.1 kHz rate works out, the stopband edge can sit at 22.05 kHz, if every alias is to be suppressed, or at 24.1 kHz, if only the audio band must stay clean. To achieve 98 dB of alias suppression — matching the 16-bit dynamic range — on the strict reading, the filter must attenuate signals at 22.05 kHz by 98 dB relative to signals at 20 kHz, across a transition band only 2.05 kHz wide. With the passband flat to within 0.1 dB, the minimum orders of the classical analogue prototypes are 13 (elliptic), 31 (Chebyshev type I) and 135 (Butterworth), or 12, 22 and 71 with the stopband at 24.1 kHz; the 44.1 kHz post’s 96 dB target gives 13, 31 and 133 at 22.05 kHz. A seventh-order elliptic filter with the same ripple reaches only about 40 dB at 22.05 kHz.
Such filters also have significant phase distortion near the passband edge: the 13th-order elliptic prototype delays 1 kHz by 58 µs and 19 kHz by 312 µs. In 1982, building such filters precisely, cheaply, and repeatably — including the reconstruction filter that every CD player needs on playback — was a genuine engineering challenge, and their phase and amplitude ripple became a suspect in early criticisms of the CD sound. I have not found a controlled listening test that settled whether it was audible.
Oversampling
Audio converters adopted the delta-sigma ($\Sigma\Delta$) architecture because it takes the steep filter out of the analogue domain, and its spread through consumer audio from the late 1980s onwards largely resolved the anti-aliasing filter debate [[1]]. The core idea is oversampling: instead of sampling at 44.1 kHz with 16 bits, the $\Sigma\Delta$ converter samples at $M \times 44.1$ kHz — where $M$ is the oversampling ratio, for example 64, giving $64 \times 44.1 = 2.8224$ MHz — with a 1-bit quantiser. By the same folding argument, the analogue filter now only has to suppress content within 20 kHz of 2.8224 MHz and its multiples; the digital decimation filter removes the rest. Its transition band runs from 20 kHz to 1.4112 MHz on the strict reading, a ratio of roughly 70:1, and to 2.8024 MHz (2.8224 MHz − 20 kHz), 140:1, on the audio-band one.
That is why a simple RC filter is acceptable, though not because it reaches 98 dB: a first-order section flat to within 0.1 dB at 20 kHz attenuates 2.8024 MHz by only 27 dB. The 98 dB figure assumed full-scale content at the stopband edge; in a well-behaved audio chain, the input at megahertz frequencies carries noise, not signal. Cirrus Logic’s data sheet for its CS5381 converter states that there is “no rejection for input signals which are (n × 6.144 MHz) ± the digital passband frequency” and recommends an input buffer with simple RC sections that “will attenuate any noise energy at 6.144 MHz” [[10]]. The steep filter has not disappeared. It has moved into the digital decimation filter, where it is exactly repeatable and can be made linear-phase.
The price paid is that the quantiser is now only 1 bit, and a 1-bit quantiser has terrible resolution on its own.
To understand what oversampling buys even before any clever signal processing, consider the quantisation noise floor. For a uniform quantiser with step size $\Delta$, the quantisation noise power is $P_q = \Delta^2/12$, and this noise is spread approximately uniformly from 0 to $f_s/2$. The noise power spectral density is $P_q / (f_s/2)$. After oversampling by a factor of $M$ — so that the effective Nyquist band runs from 0 to $f_{\text{audio}} = f_s/(2M)$ — the in-band noise power is:
$$P_{\text{in-band}} = \frac{P_q}{f_s/2} \cdot f_{\text{audio}} = \frac{P_q}{f_s/2} \cdot \frac{f_s}{2M} = \frac{P_q}{M}$$Each doubling of $M$ halves the in-band noise power, an improvement of 3 dB, equivalent to half a bit of resolution. At 64× oversampling this gives 18 dB, or three extra bits — useful, but not enough to get from a 1-bit quantiser to 16-bit performance. We need something more.
Noise Shaping
The second ingredient — and the one that makes $\Sigma\Delta$ conversion genuinely remarkable — is noise shaping. Rather than spreading quantisation noise uniformly in frequency, we can engineer its spectral distribution so that almost all the noise power sits above the audio band, where it is removed by a digital low-pass filter (the decimation filter) at the output.
A first-order $\Sigma\Delta$ modulator achieves this by a feedback loop. At each sample step, the quantiser takes the running integral of the difference between the input signal and the previous quantised output:
$$u_n = u_{n-1} + x_n - y_{n-1}, \qquad y_n = u_n + e_n,$$where $x_n$ is the input sample, $u_n$ the integrator state, $y_n = \pm 1$ the 1-bit output and $e_n = y_n - u_n$ the quantisation error. This is the integrator-feedback structure that gives the modulator its name: $\Sigma$ for the integrating summation, $\Delta$ for the difference. Substituting $y_{n-1} = u_{n-1} + e_{n-1}$ into the first equation gives $u_n = x_n - e_{n-1}$, and so
$$y_n = x_n + e_n - e_{n-1}.$$The signal passes unchanged; the error appears only as its first difference. In the $z$-domain, the quantisation noise has the transfer function
$$N(z) = 1 - z^{-1}.$$The standard analysis treats $e_n$ as white noise of power $\Delta^2/12$, independent of the signal, a bold approximation for a 1-bit quantiser. In the frequency domain, substituting $z = e^{j 2\pi f / f_s}$:
$$\bigl|N(f)\bigr|^2 = \left|1 - e^{-j 2\pi f / f_s}\right|^2 = 4\sin^2\!\left(\frac{\pi f}{f_s}\right)$$For frequencies well below the sampling rate, $f \ll f_s$, the small-angle approximation gives:
$$\bigl|N(f)\bigr|^2 \approx \left(\frac{2\pi f}{f_s}\right)^2$$The noise power spectral density rises as $f^2$ — it is heavily suppressed at low frequencies and pushed up toward $f_s/2$. A second-order modulator applies the noise-shaping filter twice, and an $L$-th order one $L$ times: $N(z) = (1 - z^{-1})^L$, with $|N(f)|^2 \approx (2\pi f/f_s)^{2L}$ at low frequencies (Fig. 2, top). Integrating the shaped noise over the audio band with the same flat density $P_q/(f_s/2)$ as before,
$$P_{\text{in-band}} = \frac{P_q}{f_s/2}\int_0^{f_s/2M}\left(\frac{2\pi f}{f_s}\right)^{2L} df = P_q\,\frac{\pi^{2L}}{(2L+1)\,M^{2L+1}}. \tag{2}$$For $L = 0$ this is the $P_q/M$ of plain oversampling. For a full-scale sine and a $b$-bit quantiser the signal-to-quantisation-noise ratio becomes
$$\text{SQNR} \approx 6.02b + 1.76 - 10\log_{10}\frac{\pi^{2L}}{2L+1} + (2L+1)\,10\log_{10} M \ \text{dB}, \tag{3}$$and an SQNR converts to effective bits as $(\text{SQNR} - 1.76)/6.02$. Each doubling of $M$ is worth $3(2L+1)$ dB: 9 dB, or 1.5 bits, for a first-order loop, and 15 dB for a second-order one. The third term is the price of shaping: the share of noise left in the band is larger than the $M^{2L+1}$ scaling alone suggests, by 5.2 dB at first order and 12.9 dB at second. At 64× oversampling — six doublings — a first-order modulator gains 49 dB over a Nyquist-rate quantiser with the same step, and a second-order one 77 dB (Fig. 2, bottom); a gain converts to bits by dividing by 6.02: 8.1 and 12.9 bits. In an earlier version of this post I gave 54 dB and 90 dB: I had kept the scaling and dropped the third term.
The following table gives the noise reduction of Eq. (3) at practical oversampling ratios:
| Modulator order | Oversampling ratio | Noise reduction, Eq. (3) | Effective bits gained |
|---|---|---|---|
| 0 (oversampling only) | 64× | 18 dB | 3.0 |
| 1st order | 64× | 49 dB | 8.1 |
| 2nd order | 64× | 77 dB | 12.9 |
| 5th order, $(1 - z^{-1})^5$ | 128× | 192 dB | 32 |
These are linear-model numbers, and a 1-bit quantiser strains the model in both directions. For a second-order loop at 64×, 3 dB below its ±1 feedback levels, the white-noise model with step $\Delta = 2$ predicts 76.2 dB; my simulation (error-feedback form, $2^{17}$ samples, Blackman–Harris window) gives 70.5 dB.
The last row is what the formula says before physics intervenes. With a 1-bit quantiser a pure third- or fifth-order loop is unstable; in the same simulation both diverge even 60 dB below full scale. Practical high-order modulators cap the out-of-band gain of the noise transfer function and pay for stability with in-band noise. A fifth-order 1-bit loop of that kind at 128× (zeros at DC, Butterworth poles, out-of-band gain 1.5) peaks in my simulation at 133 dB (about 22 bits, and 7 dB above the white-noise model) with the input 5 dB below full scale, and diverges above that.
Real converters land well below even that. Texas Instruments’ PCM1802 follows the recipe described here, “five integrators using the switched capacitor technique and a comparator” with a 1-bit DAC at 64× or 128× oversampling, and specifies a dynamic range of 105 dB typical (100 dB minimum), A-weighted, at 128× and 44.1 kHz, in 24-bit words: about 17 effective bits [[11]]. Cirrus Logic’s CS5381, a fifth-order loop with a multi-bit quantiser sampling at 6.144 MHz, specifies 120 dB typical (114 dB minimum) A-weighted and 117 dB (111 dB) unweighted, measured with a −60 dBFS signal by the AES17 method: about 19 to 20 bits [[10]]. Unless their loops are far worse than the simulated one, the gap is not quantisation noise but, presumably, the analogue circuitry: noise that enters at the input passes through the loop exactly as the signal does, and no shaping removes it.
That the PCM1802 gets about 17 bits out of a comparator — a device that outputs only 1 or 0, with no analogue subtlety whatsoever — and a few switched-capacitor integrators on a cheap CMOS chip is, I think, one of the more quietly stunning pieces of engineering in consumer electronics, and it goes entirely unnoticed because the recorder or audio interface it lives inside is now considered mundane.
Managing stability, and the limit cycles and tonal artefacts that come with losing it, is a significant part of the design problem and involves either restricting the input range, adding nonlinear stability control, or using multi-bit internal quantisers (which reduce the quantisation step and ease the stability constraint while retaining most of the noise-shaping benefit). The multi-bit approach also addresses a related issue: the ideal 1-bit DAC in the feedback loop is inherently linear (there are only two levels, so there is no differential nonlinearity), but multi-bit internal DACs must be trimmed or calibrated to avoid nonlinearity in the feedback path corrupting the noise shaping. These engineering details are discussed thoroughly in Norsworthy, Schreier, and Temes [[5]], a standard reference.
The digital audio infrastructure that delta-sigma conversion enabled — clean, cheap, phase-linear converters without steep analogue filters — also made digital audio workable in latency-sensitive applications like live performance. For a discussion of why latency matters so much in network music performance and how it shapes system design, see my earlier post on NMP latency and the physics of musical timing.
The Irony of the Comparison
Both tape saturation and delta-sigma conversion are, at root, about the same problem: how to manage the relationship between a signal and the finite resolution of the medium storing it. Tape manages the problem physically and somewhat accidentally — the ferromagnetic B-H curve happens to generate low-order harmonics that are consonant with the recorded signal, and the bias trick linearises the response well enough that the distortion only becomes audible when the engineer deliberately pushes into saturation. Delta-sigma manages the problem mathematically and deliberately — quantisation noise is redistributed in frequency by a designed feedback loop so that it falls outside the audible band.
Neither approach is perfect, and neither is neutral. Tape adds signal-correlated harmonic distortion whose spectral content depends on recording level and which compresses transients in a way that changes the perceived dynamics. Digital audio, even with delta-sigma conversion, has its own imperfections: idle-channel noise from the modulator, potential for tonal limit-cycle artefacts at specific input levels, and the abrupt onset of hard clipping at full scale — which, unlike tape saturation, is rapid and adds a long series of high-order harmonics at once, giving the harsh, unpleasant character that digital overloads are known for. In Fig. 1, a hard clip driven to twice its ceiling leaves the fifteenth harmonic about 50 dB above where the tanh curve puts it. The soft-knee vs. hard-clip distinction is real and audible, and it is probably the most defensible technical basis for the claim that analogue tape handles transient overloads more graciously.
What is not defensible is the claim that one medium is inherently more musical than the other, or that digital audio lacks something fundamental that tape possesses. They are differently imperfect. The imperfections of tape happen to sit at harmonic relationships that Western ears, shaped by a tradition of music built on those same harmonic intervals, find pleasing. The imperfections of digital audio are not at pleasing harmonic intervals; they are wideband quantisation noise (before shaping) or ultrasonic shaped noise (after), and a sharp cliff at full scale. Different physics, different perceptual character.
A Personal Note
I spent a long time thinking the tape versus digital debate was mostly audiophile mythology — a community of enthusiasts rationalising the warmth of nostalgia as the warmth of oxide particles. The physics is more interesting than that, and doing the calculation changed my view. The harmonic content of tape saturation is not an accident or a romantic story; it is what you get when you push a smooth, saturating nonlinearity with an audio sine wave (odd harmonics if it is symmetric, even ones if bias or the signal chain breaks the symmetry), and the reason it sounds pleasant is not arbitrary but is grounded in the physics of consonance and the harmonic series. The delta-sigma converter is not a mundane commodity chip but a genuinely elegant solution to an otherwise intractable filter-design problem, and the fact that it gets about 17 bits out of a 1-bit comparator by spectral redistribution of noise is the kind of result that should get more attention in physics education.
Both technologies deserve better than the aesthetics argument they have been fighting in for forty years. The tools to understand them are not exotic — Taylor series, Fourier analysis, the z-transform, and the basic physics of ferromagnetism — and the reward is a clear-eyed picture of what is actually going on inside two of the most consequential inventions in the history of recorded music. If you are interested in related mathematics underlying other aspects of music, the posts on Euclidean rhythms and Messiaen’s modes and group theory cover the combinatorial and algebraic structures in rhythm and pitch that sit alongside the physics discussed here.
References
Literature, data sheets and calculations checked through 2026-10-24.
[1] Candy, J. C., & Temes, G. C. (Eds.). (1992). Oversampling Delta-Sigma Data Converters: Theory, Design, and Simulation. IEEE Press.
[2] Reiss, J. D., & McPherson, A. (2015). Audio Effects: Theory, Implementation and Application. CRC Press.
[3] Bertram, H. N. (1994). Theory of Magnetic Recording. Cambridge University Press.
[4] Pohlmann, K. C. (2010). Principles of Digital Audio (6th ed.). McGraw-Hill.
[5] Norsworthy, S. R., Schreier, R., & Temes, G. C. (Eds.). (1997). Delta-Sigma Data Converters: Theory, Design, and Simulation. IEEE Press.
[6] Westmijze, W. K. (1956). The fundamentals of magnetic recording. Tijdschrift van het Nederlands Radiogenootschap, 21(1), 1–16. Lecture of 17 June 1955; scan by the Max Planck Institute for the History of Science. PDF
[7] Engel, F. K. (2006). Walter Weber’s technical innovation at the Reichs-Rundfunk-Gesellschaft (English version, August 2006). PDF
[8] O’Kelly, T. (n.d.). Bias. The Inventor’s Notebook, Technical Bulletin #3. BASF. Manufacturer bulletin. PDF
[9] Recording The Masters. (2022). RTM LPR35 technical data. Manufacturer data sheet. PDF
[10] Cirrus Logic. (2021). CS5381: 120 dB, 192 kHz, multi-bit audio A/D converter (data sheet DS563F3). PDF
[11] Texas Instruments. (2016). PCM1802 single-ended analog-input 24-bit, 96-kHz stereo A/D converter (data sheet SLES023D, December 2001, revised December 2016). PDF
Changelog
2026-01-14: Updated the interval description for the 7th harmonic to “above the double octave.” The 7th harmonic (7f) sits between the double octave (4f) and the triple octave (8f).
2026-07-11: Corrected the harmonic-symmetry argument: symmetric tape saturation produces odd-order, not even-order, distortion; even-order components require asymmetry. Qualified the analogue-filter, 1-bit, and effective-resolution claims for delta-sigma converters.
2026-10-03: Removed leftover even-harmonic claims that contradicted the 2026-07-11 correction (a symmetric saturating nonlinearity gives odd harmonics; even ones need asymmetry), and corrected the description of digital hard clipping, which is symmetric and therefore adds a long series of high-order harmonics rather than “all” harmonics.
2026-10-24: Corrected the noise-shaping law, which had dropped the factor $\pi^{2L}/(2L+1)$: at 64× the gains are 49 dB (8.1 bits) for first order and 77 dB (12.9 bits) for second order, not 54 dB (9 bits) and 90 dB (15 bits). Replaced the fifth-order row (~120 dB, ~20 bits) and the unsourced “fifth- to seventh-order modulators at 128× or higher” with the formula’s unrealisable 192 dB, a simulated stable fifth-order 1-bit loop (133 dB) and the typical figures of two data sheets (PCM1802: 1-bit, fifth order, 105 dB; CS5381: multi-bit, 120 dB A-weighted); “24-bit resolution from a 1-bit comparator” is now “about 17 bits”. Rewrote the first-order loop, which mixed the error-feedback and integrator forms, and replaced “was developed to take the steep filter out of the analogue domain” with why audio converters adopted it: the architecture goes back to Inose, Yasuda and Murakami’s 1962 telemetry paper, not to audio. Softened “remains the standard reference” to “a standard reference”.
Corrected the anti-aliasing filter from “typically seventh-order” to the minimum orders for 0.1 dB ripple and 98 dB (13 elliptic, 31 Chebyshev I, 135 Butterworth at 22.05 kHz; 12, 22, 71 at 24.1 kHz), and explained why an RC filter suffices for an oversampling converter although it gives only about 27 dB at 2.8 MHz. Replaced the claim that the early filters “introduced audible phase and amplitude ripple” contributing to criticism of the CD sound with “a suspect”, since I found no controlled listening test. Removed “matches the dynamic range of the best analogue tape”, which no source here supports.
On tape: corrected the Taylor collection “through fourth order”, which kept only three terms; replaced the description of AC bias as a time average with Westmijze’s 1955 account and sourced its history to Engel; replaced the unsourced 50–150 kHz range with a BASF rule of thumb; corrected the playback voltage, which follows the rate of change of flux and rises 6 dB per octave rather than reproducing the current waveform directly; changed “remaining imperfections are mostly second-order” to “of other kinds” and added the LPR35’s 0.44 per cent third harmonic; changed “musically transparent dynamic compressor” to an instantaneous limiter. Added the tanh model, two figures and references 6–11.