Why Are Scalograms So Smooth? A Comprehensive Comparison of Real and Complex Wavelets

In signal processing, the smoothness of a scalogram is often seen as an intuitive indicator of wavelet analysis quality. But why do some scalograms look smooth and refined while others are full of oscillatory texture? A recent popular Reddit discussion about wavelets explained this from the perspective of phase and amplitude, and provided a systematic comparison between real and complex wavelets.

The Root of Scalogram Smoothness: Phase and Amplitude

The essence of a scalogram is the energy distribution of a signal on the scale-shift plane. For real-valued wavelets, the frequency-domain response is asymmetric, and positive and negative frequency components interfere with each other, causing oscillations in the energy plot. The Morlet wavelet, as a representative complex wavelet, has a frequency-domain response only on the positive-frequency side. Phase and amplitude information are fully preserved, and no cross-term interference is introduced during energy calculation, so the scalogram presents a smooth envelope.

The key to this explanation is that smoothness does not come from post-processing filtering, but is a natural result of the mathematical structure of complex wavelets. The video author points out that complex wavelets encode both amplitude and phase, while real wavelets retain only the real part. The lost phase information appears as non-smooth texture in the energy plot.

Analytic wavelets and one-sided spectra: The core reason complex wavelets can avoid positive-negative frequency interference is that they satisfy the "analyticity" condition—the frequency-domain response is nonzero only on the positive-frequency side and zero on the negative-frequency side. This property makes the coefficients of the complex wavelet transform directly constitute an analytic signal, whose magnitude is the envelope (amplitude) and whose argument is the instantaneous phase. For real wavelets, the frequency-domain response is symmetrically distributed across positive and negative frequencies, and the transform coefficients are real, so amplitude and phase cannot be directly separated; if one forcibly constructs an analytic signal, post-processing such as the Hilbert transform is required, which introduces edge effects and computational burden. Therefore, analyticity is the mathematical cornerstone of the smoothness of complex-wavelet scalograms.

Real vs. Complex Wavelets: Comparison of Characteristics and Applications

Characteristics of Real Wavelets

Real wavelets (such as Haar, Daubechies, and Symlet) are computationally efficient, have real-valued coefficients, and have low storage and transmission costs. They are good at detecting abrupt changes and discontinuities in signals, and are widely used in image compression (such as JPEG 2000) and singularity detection. However, the frequency-domain asymmetry of real wavelets makes their scalograms prone to oscillation, so they are not suitable for direct instantaneous frequency estimation.

Advantages of Complex Wavelets

Complex wavelets (such as Morlet and dual-tree complex wavelets) provide approximate shift invariance, and phase information can be used for accurate instantaneous frequency and phase analysis. The Morlet wavelet achieves a good balance between time and frequency resolution and is widely used in EEG, seismic wave, and mechanical vibration analysis. The dual-tree complex wavelet achieves an approximately analytic wavelet transform through two filter banks and performs excellently in image denoising and texture analysis.

The video also briefly reviews the history of wavelets: from the Haar wavelet to the proposal of the Morlet wavelet, and then to the engineering implementation of the dual-tree complex wavelet. The design of complex wavelets has always centered on how to preserve the phase structure of signals more completely.

Dual-Tree Complex Wavelet Transform (DT-CWT): The dual-tree complex wavelet processes the signal separately through two parallel real-wavelet filter banks (usually one low-pass and one high-pass), then combines them into approximately analytic complex coefficients. Its clever design lies in the fact that the two filter banks satisfy the half-sample delay condition, so that the combined complex wavelet has approximate shift invariance and better direction selectivity than traditional real wavelets. In engineering, DT-CWT has a redundancy of only 2x (compared with the critical sampling of traditional discrete wavelets), so it is computationally efficient and therefore widely used in image denoising, texture analysis, and motion estimation.

Consensus and Divergence from Multiple Perspectives

According to the original Reddit post and the video description, the core consensus on scalogram smoothness is that the analyticity of complex wavelets (one-sided spectrum) is the mathematical basis for smooth energy plots. This view has been corroborated in discussions across multiple signal processing communities.

However, there are differing views on whether real wavelets are "completely unsuitable" for time-frequency analysis. Some practitioners point out that real wavelets combined with appropriate post-processing (such as the Hilbert transform) can also yield approximate analytic signals, but the computational pipeline is more complex. Others believe that smoothness itself is not always an advantage—when detecting transient impacts, moderate oscillation can instead provide sharper clues for time localization.

Hilbert transform and analytic signals: The Hilbert transform is a standard tool for converting a real signal into an analytic signal. By shifting the phase of the positive-frequency components by -90° and the negative-frequency components by +90°, it constructs a complex signal containing only positive frequencies. Applying the Hilbert transform scale by scale to real wavelet transform coefficients can approximately obtain analytic wavelet coefficients, from which instantaneous amplitude and phase can then be extracted. However, this method has two limitations: first, the Hilbert transform is a global operation, so local phase estimation for non-stationary signals may be distorted; second, the finite support of real wavelets leads to boundary effects that require additional handling. Therefore, although post-processing is feasible, its pipeline complexity and accuracy still fall short of directly using complex wavelets.

Practical Selection Recommendations

Choosing real or complex wavelets depends on the task objective. If accurate instantaneous frequency, phase synchronization analysis, or smooth time-frequency energy plots are required, complex wavelets (especially Morlet) are the first choice. If computational efficiency, coefficient sparsity, or abrupt-change detection is the goal, real wavelets still have irreplaceable advantages. Dual-tree complex wavelets provide a compromise between the two, trading controllable redundancy for approximate shift invariance.

Understanding the source of scalogram smoothness helps avoid mistaking visual smoothness for analysis accuracy. The true quality of time-frequency analysis depends on whether the wavelet matches the morphological characteristics of the signal, not on the visual appearance of the energy plot.

Shift invariance and redundancy: Shift invariance means that when the signal is slightly shifted, the wavelet coefficients do not change drastically. The traditional discrete wavelet transform (DWT) lacks shift invariance due to downsampling, causing coefficients to oscillate with shifts; complex wavelets (such as the Morlet continuous transform and DT-CWT) introduce redundancy through oversampling or a dual-tree structure, thereby obtaining approximate shift invariance. Redundancy is the metric for this oversampling: the higher the redundancy, the better the shift invariance, but the greater the storage and computational cost. In practice, when choosing a wavelet, one must trade off between the need for shift invariance and computational resources—for example, real-time embedded systems may prefer low-redundancy real wavelets, while offline analysis can use high-redundancy complex wavelets.

One-line Summary

Complex wavelets preserve phase information due to their one-sided spectrum, making scalograms smooth; real wavelets are efficient but lose phase and are suitable for abrupt-change detection.

Paragraph Summary

The article explains the root of scalogram smoothness from the perspective of phase and amplitude: the frequency-domain response of complex wavelets (such as Morlet) exists only on the positive-frequency side, avoiding positive-negative frequency cross-interference, so the energy plot presents a smooth envelope; real wavelets have asymmetric frequency domains and lose phase information, causing scalogram oscillations. The article systematically compares the characteristics of real wavelets (Haar, Daubechies, etc.) and complex wavelets (Morlet, dual-tree complex wavelets): real wavelets are computationally efficient and suitable for abrupt-change detection; complex wavelets provide shift invariance and accurate phase analysis and are widely used in biomedical and vibration signals. The discussion also addresses community disagreement—real wavelets combined with the Hilbert transform can approximate analytic signals, and smoothness is not always an advantage. Finally, it gives practical selection recommendations: weigh computational efficiency, phase accuracy, and smoothness according to task requirements.