What Is Multiresolution Analysis?

Multiresolution analysis (MRA) is a mathematical framework for representing and analyzing signals at multiple levels of resolution simultaneously, with each level capturing different frequency content or spatial detail. In an MRA, a signal is decomposed into a hierarchy of approximations, each successive level providing a coarser representation obtained by discarding fine-scale detail, and a corresponding set of detail coefficients that record what was removed. The framework was formalized by Stéphane Mallat in a 1989 paper in IEEE Transactions on Pattern Analysis and Machine Intelligence, which established the algebraic conditions a set of scaling functions and wavelets must satisfy for the hierarchy to be consistent and complete.

Multiresolution analysis sits at the intersection of functional analysis, digital filter theory, and applied signal processing. It provides the theoretical foundation for the discrete wavelet transform, unifying results from approximation theory and digital filter design into a single coherent structure.

Wavelet Representation and Scaling Functions

An MRA is built from two complementary functions: a scaling function (also called a father wavelet) that generates approximation spaces, and a wavelet function (mother wavelet) that generates the detail spaces between consecutive approximation levels. The scaling function satisfies a two-scale relation, meaning it can be expressed as a weighted sum of scaled and translated copies of itself, and this relation directly defines a finite impulse response low-pass filter. The wavelet function is derived from the scaling function via a high-pass filter that is the quadrature mirror of the low-pass filter. The orthogonality or biorthogonality of the wavelet basis determines the properties of the transform, including perfect reconstruction, linear phase, and compact support. Mallat's original paper, available through IEEE Xplore as a foundational signal decomposition reference, established these conditions rigorously.

Mallat's Pyramid Algorithm

Mallat also showed that computing the wavelet representation of a discrete signal reduces to iterated convolution with the low-pass and high-pass filter pair, followed by downsampling by a factor of two, a procedure now called the fast wavelet transform or Mallat's pyramid algorithm. At each level, the low-pass subband is further decomposed, producing a tree of subbands whose leaves contain the finest-scale detail coefficients. The computational complexity of this algorithm is O(N) for a signal of length N, making it significantly more efficient than the fast Fourier transform for many applications. Reconstruction proceeds by upsampling and convolving with synthesis filters, perfectly recovering the original signal when the analysis and synthesis filter banks satisfy the perfect reconstruction property. A freely available derivation of the pyramid algorithm and its filter bank structure appears in the original published paper.

Extensions and Discrete Variants

Beyond the standard one-dimensional orthonormal case, multiresolution analysis has been extended in several directions. Two-dimensional MRA applied to images decomposes a frame into four subbands at each level: one approximation subband and three detail subbands capturing horizontal, vertical, and diagonal edges. Wavelet packets generalize the tree decomposition by allowing both the approximation and detail subbands to be further split, producing a richer set of basis functions. Dual-tree complex wavelets and contourlets address the directional selectivity limitations of real-valued 2D wavelets. The theory of frames extends MRA to overcomplete representations, used in compressed sensing. Applications of these extensions to image processing and compression are surveyed in the ACM Digital Library edition of Mallat's wavelet representation paper.

Applications

Multiresolution analysis has applications in a wide range of fields, including:

• Image and video compression standards such as JPEG 2000 and motion JPEG 2000
• Seismic data processing for subsurface structure analysis
• Medical image denoising in MRI and CT reconstruction pipelines
• Audio coding and speech enhancement using subband decomposition
• Numerical solution of partial differential equations with adaptive grid refinement
• Feature extraction for texture classification and object recognition