Find the question you need
Read the open chapters from first principles and continue to the full worked solution. Use the linked chapter index to jump directly to the part you need.
Spectrum versus waveform · RMS proof · Exact first maximum · Full chapter index
Equal magnitude spectra can conceal different peaks because phase changes how harmonic components reinforce or cancel in time. This worked comparison uses two exact sine-wave sums.
Chapter index
How to distinguish a safe-looking spectrum from a trustworthy measurement
Further worked tests: sampled peaks and phase-sensitive decisions
Chapters 1–4
1. A correct spectral summary can miss an important peak · 2. Define the four quantities being compared · 3. Keep the Fourier convention consistent · 4. Construct the first signal with two sine harmonics
Chapters 5–8
5. Reverse a phase without changing its spectral magnitude · 6. Prove equal means and full-period RMS · 7. Derive Signal A’s maximum by calculus · 8. Derive Signal B’s peak
Chapters 9–12
9. Compare crest factors and threshold risk · 10. Why autocorrelation does not restore phase · 11. Sampling and missed continuous-time peaks · 12. Filters, delays and reconstruction limits
Chapters 13–14
13. Questions and further reading · 14. Practical measurement audit
Consider two hypothetical signals x_A(t)=sin(t)+sin(3t) and x_B(t)=sin(t)−sin(3t), each defined over time with a fundamental period of 2π. Both contain exactly two positive-frequency sinusoidal components, at angular frequencies 1 and 3. Both have component amplitude magnitudes of one. Yet their time-domain peaks are not the same.
The second expression reverses the third harmonic’s sign, equivalent to a change of π radians in that harmonic’s sine-reference phase. The magnitude of its Fourier coefficient is unchanged. A display that plots only positive-frequency magnitudes shows equal bars at the same two frequency locations for the two models.
That is a valuable but incomplete report: it accurately records frequency strength while omitting phase. The peak-sensitive behaviour of a waveform depends on how component crests line up in time. This worked article proves the difference with exact algebra and then connects it to RMS measurement, autocorrelation, clipping risk and sampling.
The two functions are original dimensionless teaching examples, not waveforms measured from an actual instrument. Every claim below concerns these explicit mathematical models unless a separate engineering assumption is stated.
Temporal mean is the average value of x(t) across a specified time interval. For these periodic signals we use one complete period, 0≤t≤2π. The average of either sine term is zero, so both combined signals have mean zero. This result is correct but does not imply that either waveform is always close to zero.
RMS amplitude is sqrt[(1/(2π))∫₀²πx(t)²dt]. It measures the square-root of the period-mean squared amplitude. The absolute peak is max|x(t)| over that same period. Crest factor is absolute peak divided by RMS. These definitions are related but not interchangeable.
A practical electrical or mechanical instrument may use RMS for one decision and peak for another. A voltage recorder that saturates beyond a threshold depends on instantaneous excursion. A suitable resistive energy model instead depends on the average of voltage squared. The measurement objective dictates which statistic is relevant.
When reporting values, preserve units, bandwidth, the interval, and whether the signal is a mathematical continuous-time function or a finite sampled record. A ‘maximum recorded sample’ is not automatically the maximum of the continuous function.
Our independent variable t uses angular frequencies 1 and 3 radians per chosen time unit. If t represents seconds, the corresponding ordinary frequencies are 1/(2π) Hz and 3/(2π) Hz. Both component functions have period 2π or a divisor of that period, so the full sum repeats over 2π.
A sine term sin(kt) may be represented using e^(ikt) and e^(−ikt) with complex coefficients. Under a standard two-sided Fourier-series convention, its two nonzero coefficients each have magnitude 1/2. A sign reversal flips their phases while preserving their magnitudes.
The positive-frequency amplitude plot considered here uses the peak sinusoidal amplitude convention, so both signals have magnitude one at frequencies 1 and 3. A one-sided RMS spectrum or power spectrum would carry other numerical scaling, but the comparison remains equal if a consistent convention is used.
The MIT Fourier Transform Properties lecture explains why complex coefficients have both magnitude and phase and why real signals produce conjugate-symmetric frequency representations. The periodised example here uses those principles without claiming to reproduce experimental material from MIT.
Signal A is x_A(t)=sin(t)+sin(3t). Each harmonic has peak amplitude one and is written with zero phase in the same sine-reference convention. Their instantaneous values depend on time and sometimes oppose one another despite the positive coefficient in front of both terms.
At t=0, both terms vanish. At t=π/2 the fundamental equals one while sin(3π/2)=−1, so the sum is zero. A viewer looking only at the fundamental’s ordinary positive crest would therefore miss Signal A’s actual maximum.
The waveform’s strongest excursion occurs at a different time where the two components are more constructively aligned. The exact location follows from a cubic polynomial in sin(t); we derive it below rather than trusting a plotted shape.
This model is entirely deterministic, so there is no uncertainty about its Fourier components or phase. Real instrument records can add noise, drift, harmonics outside the measured band and imperfect phase calibration. Those complications should not be used to obscure the simpler fact that magnitude information alone is mathematically incomplete.
Signal B is x_B(t)=sin(t)−sin(3t). Only the third-harmonic sign changes. The negative coefficient can be expressed as a positive-amplitude third harmonic with an additional phase π. The harmonic at angular frequency 1 remains unchanged.
Consequently a two-bar Fourier amplitude spectrum for B matches A’s exactly. The corresponding two-sided coefficient magnitudes at ±1 and ±3 also match. Any difference in power distribution by harmonic would have to come from a convention or measurement issue, not the exact models specified here.
At t=π/2, x_B=1−(−1)=2, whereas x_A at that same instant is zero. These two functions cannot be identical. Furthermore, their absolute global maxima differ, so they are not merely the same waveform shifted along the time axis.
A global time shift changes component phase in a coordinated way and generally preserves the set of amplitudes a periodic waveform reaches. Our selective harmonic sign change produces a different crest factor instead. This distinction is important before claiming that every phase adjustment necessarily changes peak amplitude.
Both mean values are zero because each sine term integrates to zero across an integer number of its full cycles in the interval 0 to 2π. The average does not distinguish the waveforms, although it accurately records that neither has a DC component.
Square x_A: sin²t+sin²3t+2sin(t)sin(3t). Square x_B: sin²t+sin²3t−2sin(t)sin(3t). Orthogonality gives ∫₀²πsin(t)sin(3t)dt=0, and the period average of each squared sine is 1/2.
Thus both period-average squared amplitudes equal 1/2+1/2=1. Both RMS values equal exactly one. The relative sign of the third harmonic affects the instantaneous sum but disappears in the whole-period mean-square calculation because the cross term averages to zero.
This conclusion should not be transferred automatically to short records, time-varying frequency content, uncalibrated RMS detectors or missing bands. The equality is established for exactly these two periodic functions and the same full-period averaging rule.
Use sin(3t)=3sin(t)−4sin³(t). Write s=sin(t), so −1≤s≤1. The first waveform becomes x_A=4s−4s³. This cubic expresses every possible amplitude of x_A without requiring a numerical sample grid.
Differentiating with respect to s gives 4−12s². Interior stationary points lie at s=±1/√3. At the positive point the value is 8/(3√3), or approximately 1.539601. The values at s=±1 are zero. Because the cubic is odd, the negative extremum has equal absolute magnitude.
The exact absolute peak is therefore 8/(3√3)≈1.539601. This is less than two: the maximum possible sum of the separate harmonic amplitude bounds is unattainable for this particular relative phase.
The calculation is independent of graph resolution. A plotted curve could approximate the peak but might slightly underestimate it if the plotting grid misses the exact time corresponding to s=1/√3.
Use the same triple-angle identity to write x_B=sin(t)−sin(3t)=4s³−2s, where s=sin(t) lies between minus one and one. At s=1, corresponding to t=π/2, this polynomial evaluates to two. Thus the absolute peak is at least two.
The triangle inequality gives a global bound for the original waveform: |sin(t)−sin(3t)|≤|sin(t)|+|sin(3t)|≤2. The exhibited value at π/2 attains that upper bound, proving the global absolute maximum equals two.
Signal A’s previously derived maximum is 8/(3√3)≈1.539601. Signal B’s is exactly two. The ratio is approximately 1.2990, even though both contain precisely the same harmonic magnitudes and have RMS one.
A proof by a global inequality and an attained equality is stronger than reading the apparent height of a curve. A digital plot might undersample a narrow peak; the analytic result is valid at every real time in the period.
Crest factor is peak absolute amplitude divided by RMS. The RMS of each signal is one, so their crest factors are 8/(3√3)≈1.5396 and exactly 2.0000. The phase-reversed signal has the greater instantaneous excursion relative to its identical average squared amplitude.
Suppose a hypothetical linear recorder becomes nonlinear or clips beyond an absolute input of 1.8 arbitrary units. Signal A remains below the limit throughout the period, whereas Signal B reaches two and crosses that threshold for some time. The different outcome does not contradict their identical RMS.
The threshold is invented as a teaching device and is not the rating of any recorder. A real device has calibration limits, analogue bandwidth, protection circuits and noise. Its response might depend on duration above threshold, not only maximum amplitude.
In a process governed by long-run mean square under a suitable linear model, equal RMS may be exactly the relevant equivalence. An instrument selection or safety claim involving peak response requires more information. The Gold Standard is choosing the metric that matches the decision.
| Property | Signal A | Signal B |
|---|---|---|
| Equation | sin t + sin 3t | sin t − sin 3t |
| Magnitude at 1 and 3 | 1, 1 | 1, 1 |
| Full-period RMS | 1 | 1 |
| Exact peak magnitude | 8/(3√3) | 2 |
| Numeric peak | 1.539601 | 2.000000 |
Define the period-mean autocorrelation as R_x(τ)=(1/(2π))∫₀²πx(t)x(t+τ)dt. Orthogonality of distinct integer harmonic orders removes the cross terms over a complete period.
Each unit-amplitude sine harmonic contributes (1/2)cos(kτ), where k is its harmonic index. The sign reversal of the third harmonic disappears when that harmonic’s autocorrelation is calculated. Hence both signals have R_x(τ)=0.5cosτ+0.5cos3τ.
This shows that even an equal second-order time statistic need not reconstruct the peak or relative harmonic phase. The two functions share the power spectrum and the periodic autocorrelation while differing in their exact instantaneous extrema.
For a real finite record, estimated autocorrelations can differ because of noise, time-window selection or imperfect sampling. That is an empirical issue, not a counterexample to the theoretical identity. Be clear whether an article reports an exact idealisation or an estimate from observed data.
Signal A reaches its maximum at times where sin(t)=1/√3. A uniformly spaced set of digital sample times need not contain any of those points. The maximum stored sample amplitude can therefore be lower than the continuous-time true maximum.
Signal B conveniently reaches its maximum at t=π/2. A grid that includes that time will record the exact value two, even though the same sampling interval may miss Signal A’s true continuous-time crest. Comparing only sampled maxima may thus introduce an additional observation-grid bias.
A suitable sample rate and reconstruction model can recover an ideally band-limited continuous waveform in principle, but exceeding a Nyquist rate is not a guarantee that one of the measured discrete samples lands precisely at a continuous extremum.
Acquisition bandwidth, clock precision, interpolation assumptions, aliasing and front-end clipping may all matter. The connected Waves guide on testing aliases explores why changing observation cadence is one way to test a suspect frequency interpretation.
A linear time-invariant filter may change both the magnitude and the phase of a frequency component. Two filters with matching magnitude responses can alter the relative phases of harmonics differently and therefore produce different time histories for the same multi-frequency input.
Not every phase change modifies the absolute peak. A pure time delay shifts the entire waveform along the time axis and preserves its amplitude range. In Fourier terms it applies a coordinated frequency-dependent phase shift that corresponds to a common time translation.
Our comparison changes the relative third-harmonic phase so the waveform shape and peak genuinely change. It is not equivalent to merely choosing another origin for time, because a common time shift could not alter the maximum absolute amplitude.
Some inverse methods estimate missing phase by adding measurements or constraints. They do not make magnitude alone a universally unique description. The two analytic functions in this article prove non-uniqueness without relying on a noisy simulation or a special device.
Do matching Fourier magnitudes guarantee matching RMS? Under the same full-band, full-period conventions in our deterministic example they do, consistent with Parseval-type relations. That does not imply matching maximum.
Could the two signals be the same waveform shifted in time? No. A pure shift preserves the range of amplitudes and therefore the absolute peak. Their distinct maxima rule that out.
Does a phase reversal always change crest factor? No. The result depends on the combination of harmonics and initial phases. Our exact pair shows one case where it does, rather than a universal rule.
What if the third-harmonic coefficient is scaled to one-half? Write x(t)=sin(t)+0.5sin(3t). The RMS becomes sqrt(1/2+0.5²/2)=sqrt(0.625), and the peak must be recalculated. This is a changed magnitude spectrum, not merely a changed phase.
Practice with x_φ(t)=sin(t)+sin(3t+φ). Its full-period mean-square remains one for any fixed φ because the cross terms vanish, while the instantaneous maxima vary with φ. Check φ=0 and φ=π against the two exact results before exploring intermediate angles.
A spectrum should identify whether its vertical axis shows amplitude, RMS amplitude, magnitude-squared or a density per hertz. It should state one-sided or two-sided scaling, frequency units, sampling interval and window function. A precise-looking plot can be quantitatively misleading without those definitions.
Peak-sensitive conclusions also require retained phase or suitably calibrated time-domain data, adequate bandwidth and explicit uncertainty. A frequency-magnitude representation cannot be relied upon to reproduce instantaneous clipping, impulse behaviour or time-localised waveform detail without additional assumptions.
To study related concepts, use The Gold Standard of Signal Processing, The Gold Standard of Spectral Analysis, Electrical Engineering: Equal RMS and Different Crest Factors, and Acoustics: Logarithmic and Energy Averages.
The MIT Fourier Properties lecture explains the theory of complex coefficients, phase, linearity and Parseval’s relation. Our two sine sums and illustrative clipping limit were constructed independently as a transparent pedagogical example.
Return to the Gold Standard hub or How X Works. Ask which part of the data the intended outcome depends on before relying on a convenient summary.
Advanced worked companion: phase invariance and missing-data checks
To make the phase dependence explicit, consider y_φ(t)=sin(t)+sin(3t+φ). Every fixed φ preserves the harmonic-amplitude magnitudes: the first and third harmonics each have unit amplitude. The RMS over one complete 2π period also remains exactly one, because the period average of each squared harmonic is one-half and the cross term between distinct harmonic indices vanishes for every constant phase.
This does not mean the waveform’s peak is fixed. Our special cases φ=0 and φ=π produce exact peak magnitudes 8/(3√3) and two. At a general φ the maximum can be found by differentiating dy_φ/dt=cos(t)+3cos(3t+φ), identifying candidate times and checking their values. A numerical routine should search the entire declared interval rather than trusting an unlabelled plot or the first local stationary point it encounters.
Phase estimation can become ambiguous when a record is very short, when an instrument has an unknown trigger delay or when frequency bins contain leakage from neighbouring tones. A perfectly resolved deterministic Fourier coefficient has a mathematically precise phase; an estimated coefficient carries sampling and calibration uncertainty. Do not report an empirical phase angle to unjustified precision when the original data did not support it.
RMS alone also does not identify the duration of a high excursion. A waveform could have a sharp but brief peak while another has a flatter response over a longer interval. If a physical outcome depends on accumulated time above a limit, integrate an explicit indicator or physical response model over the interval rather than substituting the instantaneous maximum or RMS indiscriminately.
For an arbitrary scaling c, replacing both x_A and x_B by c times their values multiplies RMS and peak magnitude by |c| but leaves crest factor unchanged. Scaling does not repair the missing phase problem. The two signals still have distinct crest factors after any nonzero common scaling.
A robust analysis can therefore report a compact bundle: mean, RMS, peak, crest factor and—where reconstruction matters—phase-preserving frequency coefficients or a calibrated waveform. Each member has a distinct meaning, and readers should know exactly which components support the claimed threshold or performance result.
Return to chapter index · Signal Processing subject root
Further worked tests: sampled peaks and phase-sensitive decisions
Take a sampling grid of eight equally spaced points per fundamental period, at t=kπ/4 for k=0 through 7. It includes π/2, where Signal B reaches its exact maximum two. It does not include a point where sin(t)=1/√3, the condition for Signal A’s exact maximum. These two sample records are therefore not equally representative of the continuous peaks even though the sampling times are identical.
On that eight-point grid, Signal A reaches a largest absolute sampled value of √2≈1.414214 at t=π/4 and at several symmetry-related times. Its continuous-time peak is 8/(3√3)≈1.539601. The sampled maximum is lower by about 0.125387 amplitude units, or roughly 8.14 percent of the true maximum. Signal B happens to have its maximum located on the grid.
The result distinguishes a discrete-data maximum from a mathematical continuous-time maximum. A waveform reconstruction method with justified band-limited assumptions may locate peaks between samples; simply viewing the stored numbers cannot. Reporting only the largest stored sample can conceal an actual excursion through a physical threshold.
The theoretical Fourier-series coefficients provide another independent check. Under a two-sided convention, sin t has nonzero coefficients of magnitude one-half at frequency indices ±1; sin3t has the same magnitude at ±3. The four squared magnitudes sum to 4×(1/2)²=1, consistent with the calculated full-period mean square. Reversing the third harmonic changes phase but not these squared magnitudes.
An apparent disagreement between a Fourier coefficient plot and a time-domain RMS should therefore first prompt a convention audit: one-sided versus two-sided normalisation, peak versus RMS amplitude, sample count scaling, spectral leakage and incomplete bandwidth. A correct Fourier identity does not rescue a frequency plot that was scaled inconsistently.
For any nonzero common amplitude scaling c, both signal peaks and RMS levels scale by |c|, while the crest factor remains unchanged. This is why the comparison is about waveform shape rather than a trivial choice of units. If only the third-harmonic amplitude changes, the frequency magnitudes no longer agree, and the simple phase-only counterexample has changed.
Measurement planning should separate at least three evidence requirements: average square for energy-like responses, exact or bounded peak for nonlinear threshold responses, and phase or time-history information for reconstructing waveform shape. No one of these measurements is a universal substitute for all others.
Finally, a numerical FFT on a finite noisy data record estimates frequency-domain quantities under windowing assumptions. Its returned phase may be noisy or reference-time dependent. A high-confidence engineering conclusion should state the sampling clock, calibration, frequency support and uncertainty alongside any claimed crest factor or threshold exceedance.
Return to chapter index · Gold Standard hub
How to distinguish a safe-looking spectrum from a trustworthy measurement
A frequency plot is not a complete record of a physical event unless the purpose of the record is explicitly limited to frequency-magnitude content. For instance, a device may be sensitive to a brief overvoltage, a structural component to a peak dynamic displacement, and a thermal load to an average squared signal. Those decisions use different operations on data and may require additional physical models beyond Fourier analysis.
In an audit, the reviewer should compare the analogue input range with the maximum possible excursion before checking any digital FFT result. If a front-end amplifier saturated, a later spectrum may accurately describe the clipped waveform while concealing what the input would have been without clipping. The mathematical phase counterexample is a reminder to inspect how data were acquired, not just how a final plot was calculated.
Another important check is whether all relevant harmonics were captured. The exact equality between our two signals is confined to two known sinusoids with no missing components. Unknown higher-frequency components may affect continuous-time peaks even if they contribute little to a low-bandwidth RMS reading. A band limit is a model or instrument property that requires evidence.
When selecting a summary, state whether it supports an average energy estimate, spectral identification, timed event detection or peak-bound guarantee. If a guarantee concerns a maximum, use sufficient time-domain evidence or a mathematically justified bound. Do not infer peak safety from a matching amplitude spectrum without phase information and measurement validation.