Two sensors can record different mixtures of the same underlying signals. Recovering the sources is straightforward if a suitable mixing matrix is known and invertible. In blind source separation the mixing is not known, so the observations alone do not automatically determine a unique answer. Statistical assumptions, normalization and the available data all matter.
A linear mixture is a model with assumptions
At a given time let be two source amplitudes and let be two sensor readings. An instantaneous linear noiseless model writes , where a fixed matrix contains the mixing coefficients. “Instantaneous” means the same time sample of each source contributes to the current observation; it excludes delayed echoes and convolution unless the state is expanded or the model changes. A noise term would give and invalidate exact recovery by simply applying an inverse.
Definition 12.1.1
(Instantaneous linear mixture)
Let be the source vector and the sensor vector at one time sample. The instantaneous noiseless linear mixture model is , where the fixed square matrix has entry equal to the weight of source in sensor . In this unit in the numerical examples.
If is known and , then uniquely recovers the source vector at each time. If is singular, there is a nonzero vector with ; and have the same observation. Thus a single observation cannot distinguish those two possible source vectors without further restrictions. Rank, sensor count and source assumptions are part of the model, not details that an algorithm can ignore.
Proposition 12.1.2
(Recovery from a known mixing matrix)
Let be a real matrix and .
- If , then is the only source vector that produces .
- If , the source vectors that produce are exactly with , and there are infinitely many of them.
Proof
(Uniqueness and the kernel)
If and , multiplying by gives , so the solution is unique. If , then has a nonzero kernel vector . For every with we have . Conversely, if then , so with . The kernel contains every real multiple , so it is infinite.
Worked problem: known invertible mixing
Worked example 12.1.3
(Recovering two sources from two readings)
Let
The equations are and . Adding and subtracting gives and . The determinant is , so the inverse exists; substituting gives as required. This is known-matrix separation, not a solution to the blind problem.
Worked problem: singular mixing has many solutions
Worked example 12.1.4
(A singular mixing leaves a whole line of solutions)
Let and . Both and produce the same . In fact every pair satisfying is a solution. The second sensor adds no independent linear equation: its row is twice the first. A particular application might impose extra constraints that select one possibility, but those constraints must be stated and justified; the matrix and observation alone do not select it.
Why blind factorization has ambiguities
Even if an unknown square mixing matrix is invertible, the factorization is not unique without restrictions. For any invertible matrix , write : a different mixing matrix and different candidate sources give the same observations. Statistical restrictions can remove many possibilities, but basic scale and order ambiguity remain for standard independent-component models. Multiplying one source by a nonzero constant and dividing the corresponding mixing column by that constant changes neither nor independence. Permuting the sources and the mixing columns together also changes neither observation nor independence.
Proposition 12.1.5
(Equivalent factorizations of a mixture)
Let with an invertible matrix, and let be any invertible matrix. Then . If, in addition, the coordinates of are independent and is a diagonal matrix with nonzero diagonal entries or a permutation matrix, then the coordinates of are independent.
Proof
(Substitution and coordinatewise maps)
Matrix multiplication is associative, so . If , the th coordinate of is , a function of alone; functions of independent random variables, each using a different variable, are independent. If is a permutation matrix, has the same coordinates as in another order, and independence of a family does not depend on how it is listed.
Therefore a separation algorithm's output can at best be compared after stating a scale, sign and permutation convention, unless extra information fixes them. A recovered component is not automatically labeled “speaker 1” or guaranteed to have the original physical units. Correlation with a known reference, calibration, or application-specific information can help label or scale outputs; it is additional evidence rather than a consequence of blind factorization.
Covariance and whitening
Assume now that sources and observations are random vectors with finite second moments and mean zero. Their covariance matrices satisfy
under the noiseless model. If is positive definite, a whitening matrix satisfies . The whitened observation has identity covariance. For example, an eigendecomposition with positive eigenvalues gives . Centering is needed for a covariance-based calculation when means are not already zero. If a covariance eigenvalue is zero, this full inverse square root is undefined; discarding zero or small directions changes the problem and needs a stated rank and noise policy.
Definition 12.1.6
(Whitening matrix)
Let be a mean-zero random vector whose covariance is positive definite. A matrix is a whitening matrix for if . Then has covariance .
Theorem 12.1.7
(Whitening matrices and their rotation freedom)
Let with orthogonal and diagonal with positive entries. Then:
- is a whitening matrix.
- A square matrix is a whitening matrix if and only if for some orthogonal matrix .
Proof
(Checking the formula and the converse)
For part 1, . For part 2, if with orthogonal, then . Conversely, let and put ; is invertible because and are. Since , we get . A square matrix with is orthogonal, and .
Proof X-ray
How the converse of the whitening theorem closes
Part 1 is substitution; the work is the converse in part 2. The idea is to compare an arbitrary whitening matrix with through the ratio and show that is orthogonal. Positive definiteness is used twice: it makes exist, so is invertible and the ratio is defined, and it gives . That identity turns the hypothesis into . The proof closes because a square matrix with is orthogonal. Squareness matters: a rectangular , for example one row of , can satisfy without being for any orthogonal .
Worked problem: whiten a two-sensor covariance
Worked example 12.1.8
(Whitening a two-sensor covariance)
Use the matrix from the invertible-mixing example and suppose the zero-mean sources are uncorrelated with variances 1 and 4. Then and
The unit eigenvectors and have eigenvalues 2 and 8. One whitening matrix is
Indeed , so and its covariance is . In this specially chosen example the whitening directions happen to align with the sources after rescaling. The calculation does not prove that whitening alone separates arbitrary mixtures.
By the whitening theorem, also whitens for every orthogonal , so covariance alone leaves a rotation ambiguity. Under the additional square invertible model in which the sources have identity covariance, the whitened mixing matrix is orthogonal: . That statement does not mean the orthogonal matrix is automatically known or equal to the identity. Whitening standardizes second-order structure; it does not in general identify the source axes.
Concept lensGeometric
Whitening as flattening an ellipsoid into a sphere
whitens if and only if , that is, for every . So carries the ellipse onto the unit circle (sphere in higher dimension), and the freedom is the freedom to rotate or reflect that sphere. In the worked example has eigenvalues and , so the ellipse has semi-axes along and along .
Common mistake
(Reading the whitened output as the sources)
After whitening, has identity covariance, and it is tempting to take to be the sources, or the sources up to scale. That held in the whitening example because the example was chosen so that the covariance eigenvectors aligned with the source axes. In general , and may be any orthogonal matrix when , which mixes the sources into one another. Whitening fixes only second-order structure.
Uncorrelated is not the same as independent
Independence implies zero covariance when the relevant moments exist, but zero covariance does not generally imply independence. A rotation of whitened independent non-Gaussian components illustrates the distinction.
Counterexample mode
Identity covariance does not imply independence
It is tempting to infer independence from identity covariance. Let and be independent, each equally likely to be or , and set , . The original vector has mean zero and identity covariance, and the transformation is orthogonal, so also has identity covariance. Yet and , while . Independence would require the joint probability to be . Thus the rotated components are uncorrelated but dependent. A covariance-only test cannot distinguish all rotations of whitened data. Independence requires the joint distribution to factor, which is a statement beyond covariance.
Independent component analysis uses assumptions beyond covariance to select components that are statistically independent, or approximately so under an explicitly chosen objective. Non-Gaussianity often provides information about orientation; a spherical jointly Gaussian vector remains distributionally unchanged by orthogonal rotation, so covariance and the full Gaussian distribution cannot identify a unique orientation in that example. This observation is not a proof of every ICA identifiability theorem. Particular guarantees require their precise independence, non-Gaussianity, rank and sampling hypotheses.
From an ideal calculation to data
In real recordings the mean and covariance are estimated from finite samples. Sampling variability can make the estimated covariance inaccurate; small eigenvalues can amplify noise when inverted. A linear instantaneous model may also be inadequate for reverberant audio, where a source affects a sensor at several delays. Sensor noise, more sources than independent sensors, changing mixtures and dependent sources each require a different model or additional assumptions. A clean matrix inverse on one noiseless example does not establish robust recovery for those cases.
The order of a practical pipeline should therefore be explicit: define the observed variables and sampling units; center using a stated estimate; choose a rank and whitening convention; specify the separation criterion; and evaluate the outputs against an appropriate reference or task. If a preprocessing transform is learned for a predictive pipeline, fit it using the intended training data and apply the saved transform consistently to validation and test data. Avoid treating a visually plausible waveform or a zero sample covariance as proof of independence.
A useful check is to reconstruct the observations from estimated sources and a mixing matrix. A small reconstruction error shows that those factors reproduce the observed data under the chosen metric; it does not by itself prove that the factors are the unique true physical sources. The scale and permutation ambiguities already demonstrate that several equivalent factorizations can reconstruct exactly. State what additional evidence justifies any interpretation of a component, including its label and units.
The unit-variance convention does not fix everything
Suppose independent sources have nonzero finite variances. Rescaling each source to variance 1 and multiplying its mixing column by the old standard deviation preserves the mixture. This removes the magnitude of a scale ambiguity under that convention, but the sign of each standardized component can still be flipped, and the components can still be permuted. A statistical convention therefore selects a convenient representative from a class of equivalent models; it does not recover original sound-pressure units on its own.
For a concrete scale example, the first column of the invertible-mixing example is . Replace by and that column by . Its contribution to both observed channels is unchanged. With the same observation , the alternative source coordinates are . Both reconstructions are exact. An external calibrated measurement of source amplitude, if available and trustworthy, can distinguish physical units; without it the two factorizations are observationally equivalent.
A reconstruction check can pass for the wrong interpretation
Let in the general transformation above. It is invertible, and . Define . Then identically, so every sensor sample is reconstructed exactly. Yet the first new coordinate combines both original sources. If the original sources are independent and is finite, then . The two new coordinates are therefore dependent, since independent variables with finite second moments have zero covariance. An independence assumption can reject this transformation under these hypotheses. Reconstruction alone cannot.
Common mistake
(Taking a perfect reconstruction as identification of the sources)
A zero reconstruction error shows only that reproduces . By the equivalent-factorization proposition, reconstructs exactly for every invertible , including a that mixes the true sources. Exact reconstruction therefore cannot tell the true sources from a mixed version of them; an independence assumption can, as the example above shows.
In contrast, a permutation or nonzero separate rescaling of independent coordinates preserves their independence. The mixing factorization can therefore remain ambiguous even after independence is imposed. One should report a source-separation result with its permitted equivalences and assumptions instead of claiming that a mathematically exact inversion necessarily gives uniquely labeled physical objects.
The finite-sample version introduces another distinction: sample covariance equal to identity is a statement about a finite dataset and chosen normalization. It does not prove population covariance is exactly identity or that population components are independent. An independence criterion optimized on the same data can also overfit. Evaluation design must match the downstream claim: reconstruction error, independence diagnostics and performance on a downstream classification task answer different questions. None alone supplies every conclusion about physical source identity.
Exercises
Quick check
(Let and . Describe every source vector with .)
Compare the two rows of .
Solution · Answer
The first row gives . The second row is times the first, and gives , the same condition. Hence and the solutions are for every real , that is, plus any multiple of the kernel vector .
Quick check
(A zero-mean vector has . Give a diagonal whitening matrix and a second whitening matrix that is not diagonal.)
Use with an orthogonal .
Solution · Answer
Take : then . With the rotation , the matrix is not diagonal, and .
Quick check
(Let be uniform on and . Show that and are uncorrelated but not independent.)
Compute , , and compare with .
Solution · Answer
and , so . But forces , so , while . The variables are dependent.
Quick check
(With and , use a sign flip of the second source to give a second factorization of the same .)
Take in the equivalent-factorization proposition.
Solution · Answer
Here , so and . Then . Both factorizations give , so the sign of a recovered source is not determined by the observations.