Does JADE require labeled source examples?

Its blind-separation formulation does not.

Are output order and sign uniquely determined?

No; conventions or external information are needed.

AI Glossary / Statistics and probability / Advanced

Joint Approximation Diagonalization of Eigen-matrices

Definition

JADE, conventionally expanded as Joint Approximate Diagonalization of Eigenmatrices, is an independent component analysis method that uses fourth-order cumulant information to separate linearly mixed sources. After centering and whitening observations, it seeks a common rotation that makes a collection of cumulant-related matrices nearly diagonal. The supplied glossary title uses Approximation; Approximate is the usual methodological name.

What it means

Several sensors may each record a different mixture of the same underlying signals. JADE attempts to recover source components from statistical structure in those mixtures rather than from labeled examples of each source.
The independence assumptions are substantive. Two uncorrelated channels are not necessarily independent, and a mathematically separated component does not automatically correspond to a named physical process. Interpretation and validation remain separate tasks.

How it works

Center the observations and whiten them so their covariance becomes the identity on the retained subspace. Estimate fourth-order cumulant matrices, then search for an orthogonal rotation that reduces their combined off-diagonal entries. Joint diagonalization uses one transformation for the whole collection.
Fourth-order separation requires informative fourth-order structure. In the standard population formulation, at most one source may have zero excess kurtosis. Merely assuming that sources are non-Gaussian is insufficient for this particular condition, because a non-Gaussian distribution can also have zero excess kurtosis.
Combine the rotation with whitening to form an unmixing transformation. Estimated sources retain order and sign ambiguities; scale also requires a convention. Finite samples make exact diagonalization unlikely, hence approximate.

Three practical examples

A researcher separates simulated sensor mixtures whose independent sources and mixing matrix are known, allowing recovery to be checked directly.
A signal-processing team investigates whether several industrial measurements contain distinct independent components suitable for further analysis.
An AI feature-development team compares components obtained across datasets and checks stability before treating them as meaningful inputs.

Think of it this way

Imagine turning a collection of transparent diagrams together until the same axes make all their crosswise markings as small as possible. Rotating each diagram separately would miss the point: the common orientation is the information being sought.

Worked example

Consider an original idealized two-source example. Let source s1 take values minus one and plus one with equal probability. Its mean is zero, variance is one, and fourth moment is one, giving excess kurtosis one minus three = minus two.
Independently, let s2 be zero with probability three quarters, minus two with probability one eighth, and plus two with probability one eighth. Its mean is zero. Its second moment is 4 times one quarter = one, while its fourth moment is 16 times one quarter = four. Its excess kurtosis is therefore one.
Mix the sources using the orthogonal matrix U with rows (0.6, minus 0.8) and (0.8, 0.6). The observations are x1 = 0.6s1 minus 0.8s2 and x2 = 0.8s1 plus 0.6s2.
Because the sources are independent with unit variance and U is orthogonal, the observed covariance is already the identity. Whitening alone therefore does not undo this rotation. Covariance supplies no preferred orientation among orthogonal rotations in this example.
Let u1 and u2 denote the columns of U. Consider the two ideal cumulant-related matrices Q1 = minus 2 times u1 times its transpose and Q2 = u2 times its transpose. They represent the two nonzero source directions with their excess-kurtosis weights.
Q1 has rows (minus 0.72, minus 0.96) and (minus 0.96, minus 1.28). Q2 has rows (0.64, minus 0.48) and (minus 0.48, 0.36).
Their combined sum of squared off-diagonal entries is twice 0.96 squared plus twice 0.48 squared = 2.304. Both symmetric off-diagonal positions count in this convention.
Apply U transpose on the left and U on the right to both matrices. Q1 becomes diagonal with entries minus two and zero. Q2 becomes diagonal with entries zero and one. Their combined off-diagonal sum is now zero, the smallest possible value because it is a sum of squares.
The same inverse rotation recovers the sources: s1 = 0.6x1 plus 0.8x2 and s2 = minus 0.8x1 plus 0.6x2. For the possible source realization s1 = 1 and s2 = 2, the mixed observations are minus one and two. Substitution recovers one and two exactly.
Swapping the recovered outputs or reversing either output's sign still yields independent components. Additional labels or physical conventions are needed to decide which output should be called source one and which sign is meaningful.
Now perturb one estimated matrix slightly while leaving the other unchanged. Their preferred axes may no longer coincide exactly. A common rotation must then compromise across the collection; diagonalizing each matrix independently does not provide one consistent unmixing transformation.
This example uses specified population quantities and a known rotation to demonstrate the target operation. It is not a claim that JADE was fitted to a finite dataset or that real noisy signals will be recovered exactly.

Illustrative case study

Fictional company Willow Sensors evaluates source separation using controlled synthetic mixtures. The team checks reconstructed components against known sources after allowing for permutation and sign changes.
It then examines sensitivity to sample size, unusual observations, and deviations from the assumed mixing model. Components that appear unstable are not assigned operational labels merely because a diagonalization criterion improved. This fictional evaluation illustrates validation, not demonstrated performance on a deployed system.

Common misconceptions and mistakes

Whitening is not the same as independent component recovery.
Fourth-order cumulants are not interchangeable with raw fourth moments.
Separate diagonalizations do not guarantee a shared source basis.
A lower fitting criterion does not prove physical interpretability.
Non-Gaussianity alone does not ensure that fourth-order information identifies every component.

Business use

JADE can support exploratory signal separation when a linear mixing model and source-independence assumptions are credible. Validate recovered components against known test mixtures or independent operational evidence. Assess data quality and stability before using components in monitoring, prediction, or automated decisions.

Check your understanding

What is the observed covariance in the example?

Answer: The identity matrix.

What is the initial off-diagonal sum?

Answer: 2.304.

What happens after the common inverse rotation?

Answer: Both supplied matrices become diagonal.

Related terms

Independent component analysis
blind source separation
whitening
cumulant
excess kurtosis
orthogonal rotation
joint diagonalization
mixing matrix

Learn next

Study cumulant estimation, ICA identifiability, and the distinction between population separation and finite-sample recovery.

Sources

Jean-Francois Cardoso, Joint diagonalization: https://www2.iap.fr/users/cardoso/jointdiag.html . Miettinen, Taskinen, Nordhausen and Oja, Fourth Moments and Independent Component Analysis, Section 5.3: https://arxiv.org/html/1406.4765 . Source distributions, matrices, and company scenario are original illustrations.

Previous: Joint action Next: Joint compatibility branch and bound