Will the real Jennrich's Algorithm please stand up?

Thanks to Rasmus Bro, Nikos Sidiropoulos, Aravindan Vijayaraghavan, and Aditya Bhaskara for helpful discussions on this topic and contributions to this post.

Summary

In many papers on tensor decomposition since 2014, the simultaneous diagonalization algorithm is incorrectly referenced as Jennrich’s algorithm. This method should not be attributed to Jennrich or even Harshman (1970) but instead referred to simply as the simultaneous diagonalization method. The best single reference for the algorithm is Leurgans, Ross, and Abel (1993), though other relevant papers can also be cited as discussed further in this post.

Background

The CP tensor decomposition was first discovered by Hitchcock (1927) and then rediscovered independently in 1970 by two groups who also each proposed the same algorithm for computing it:

Harshman’s paper credits Robert Jennrich for the algorithm (and a proof of uniqueness). The algorithm that Jennrich proposed in Harshman (1970) is not Simultaneous Diagonalization but instead Alternating Least Squares, the same as that proposed by Carroll and Chang. Here is an excerpt from Harshman (1970) that specifies the algorithm: The simultaneous diagonalization algorithm wrongly credited to Jennrich has an interesting history that is discussed in detail in Appendix 6.A of A. Smilde, R. Bro, and P. Geladi (2004), Multi-Way Analysis: Applications in the Chemical Sciences, Wiley. See also the background discussion in L. De Lathauwer, B. De Moor, and J. Vandewalle (2004), Computation of the Canonical Decomposition by Means of a Simultaneous Generalized Schur Decomposition, SIAM J. Matrix Analysis and Applications, 26:295-327. (see also Further Reading below).

For these reasons, this method has no obvious progenitor for whom it should be named, and the most apropos single reference is

The “real” Jennrich’s algorithm is BETTER!

The simultaneous diagonalization only works for three-way tensors and is numerically finicky. There are robust MATLAB implementations of simultaneous diagonalization in both N-way Toolbox (nwy_dtld) and TensorLab (tlb_sd), so we compare these to the “real” Jennrich’s algorithm, which is the well-known Alternating Least Squares (ALS). We do this on a tensor of size $25 \times 20 \times 15$ with rank $r=10$ and 20% noise and correlated columns in the factor matrices (congruency of 0.8). The ALS method achieves a better fit of 0.81 versus 0.80 for simultaneous diagonalization and, more importantly, a much better correlation with the noise-free solution of 0.925 versus 0.651. (Full details are given at the end of this posting.)

A Connection with respect to Uniqueness

One point of similarity between the paper of Leurgans-Ross-Abel and Jennrich (via Harshman) is that they proved uniqueness of the CP tensor decomposition under similar circumstances.

• Jennrich (Harshman 1970): The rank-$R$ decomposition $X = [[A,B,C]]$ is unique (up to permutation and scaling) if $\text{rank}(A) = \text{rank}(B) = \text{rank}( C ) = R$.

• Leurgans-Ross-Abel (1993): The rank-$R$ decomposition $X = [[A,B,C]]$ is unique (up to permutation and scaling) if $\text{rank}(A) = \text{rank}(B) =$ and $\text{K-rank}( C ) = 2$ (i.e., any pair of columns in $C$ are linearly independent).

There are two important differences between these results.

(1) There is no connection between Jennrich’s result and the algorithm he proposed. In contrast, the simultaneous diagonalization algorithm of Leurgans-Ross-Abel was directly connected to their theory.

(2) The condition on $C$ is stricter in Harshman’s version.

A Stepping Stone

There is another paper by Harshman that derives the same result as Leurgans-Ross-Abel:

It does credit Robert Jennrich and David Cantor for constructive feedback, but the credit for this result goes to Harshman himself. So we add another uniqueness milestone:

• Harshman (1972): The rank-$R$ decomposition $X = [[A,B,C]]$ is unique (up to permutation and scaling) if $\text{rank}(A) = \text{rank}(B) =$ and $\text{K-rank}( C ) = 2$ (i.e., any pair of columns in $C$ are linearly independent).

A few comments are in order.

(1) No algorithm is provided, though the proof is constructive.

(2) Leurgans-Ross-Abel (1993) did not cite Harshman (1972), and their proof seems to have been developed independently.

An Aside on Uniqueness

As an aside, we should also note a more general result which does not come with an algorithm for its computation. The reference is

The result is

• Kruskal (1977): The rank-$R$ decomposition $X = [[A,B,C]]$ is unique (up to permutation and scaling) if $\text{K-rank}(A) + \text{K-rank}(B) + \text{K-rank}( C ) \geq 2R+2$ where K-rank is defined to be the minimum value $K$ such that any $K$ columns are linearly independent.

A nice explanation of the uniqueness conditions and algebraic solution is given here:

Other papers that are often credited in the evolution of these ideas are:

Perhaps the most general version of the algebraic solution method is provided here:

Details on Experimental Results

Setup

• We create a three-way approximately low-rank tensor with the following characteristics. Size is 25 x 20 x 15. X = 80% Xtrue + 20% Normal-Distributed Noise. Xtrue = rank-10 tensor with factors [[A,B,C]] with normalized columns and such that the dot products between any two columns of A (or B or C) is exactly 0.8. (For this we use the method described by Tomasi and Bro (2006) and implemented in the Tensor Toolbox as matrandcong.)

• We compare the ALS method tlb_als from TensorLab, with two simultaneous diagonalization (SD) methods: tlb_sd from TensorLab and nwy_dtld from N-way Toolbox

• The correlation score of the computed solution and the true solution is computed via the score function in Tensor Toolbox. A score of 1.0 is ideal, and a score of zero indicates no correlation.

• We run each method 10 times and report the best fit and correlation score of the 10 runs. Note that nwy_dtld does not have an option to set the initial guess.

Output

TIME

Method tlb_als tlb_sd nwy_dtld
Run 1 0.046 0.036 0.060
Run 2 0.018 0.023 0.019
Run 3 0.024 0.010 0.016
Run 4 0.017 0.008 0.017
Run 5 0.016 0.021 0.016
Run 6 0.012 0.010 0.016
Run 7 0.021 0.009 0.017
Run 8 0.023 0.015 0.016
Run 9 0.016 0.011 0.015
Run 10 0.019 0.010 0.016
SUM 0.212 0.153 0.206

FIT

Method tlb_als tlb_sd nwy_dtld
Run 1 0.811 0.800 0.793
Run 2 0.811 0.800 0.793
Run 3 0.808 0.797 0.793
Run 4 0.811 0.796 0.793
Run 5 0.812 0.800 0.793
Run 6 0.811 0.797 0.793
Run 7 0.811 0.796 0.793
Run 8 0.812 0.800 0.793
Run 9 0.812 0.799 0.793
Run 10 0.812 0.800 0.793
MAX 0.812 0.800 0.793

SCORE

Method tlb_als tlb_sd nwy_dtld
Run 1 0.910 0.643 0.568
Run 2 0.850 0.646 0.568
Run 3 0.740 0.544 0.568
Run 4 0.871 0.520 0.568
Run 5 0.921 0.638 0.568
Run 6 0.876 0.608 0.568
Run 7 0.896 0.605 0.568
Run 8 0.875 0.651 0.568
Run 9 0.846 0.566 0.567
Run 10 0.925 0.642 0.568
MAX 0.925 0.651 0.568

Source Code

%% Comparing ALS and SD

%% Setup path
% Be sure to add the Tensor Toolbox (Version 3.2.1), the N-Way Toolbox, and
% TensorLab (Version 3.0) to your path in that order. In the search path,
% tensorlab should be above, nway331, which is above tensor_toolbox.

%% Create a test problem
m = 25; n = 20; p = 15; r = 10;

rng('default') %<- For reproducibility, feel free to omit

info = create_problem('Size',[m n p],'Num_Factors',r,'Noise',0.2,...
'Factor_Generator', @(m,n) matrandcong(m,n,0.8), ...
'Lambda_Generator', @ones);

X = info.Data;
Mtrue = arrange(info.Soln);

%% Run experiments

rng('default') %<- For reproducibility, feel free to omit

ntrials = 10;
results = cell(ntrials,1);
for i = 1:ntrials

U1 = randn(m,r);
U2 = randn(n,r);
U3 = randn(p,r);
Minit = {U1,U2,U3};

tic
M = cpd_als(X.data,Minit);
results{i}.tlb_als.time = toc;
M = ktensor(M);
results{i}.tlb_als.fit = 1 - norm(X-full(M))/norm(X);
results{i}.tlb_als.score = score(M,Mtrue,'lambda_penalty',false);

tic
M = cpd3_sd(X.data,Minit);
results{i}.tlb_sd.time = toc;
M = ktensor(M);
results{i}.tlb_sd.fit = 1 - norm(X-full(M))/norm(X);
results{i}.tlb_sd.score = score(M,Mtrue,'lambda_penalty',false);

% No way to set the initial guess for DTLD
tic
[A,B,C] = dtld(X.data,r);
results{i}.nwy_dtld.time = toc;
M = ktensor({A,B,C});
results{i}.nwy_dtld.fit = 1 - norm(X-full(M))/norm(X);
results{i}.nwy_dtld.score = score(M,Mtrue,'lambda_penalty',false);

end

%% Print out results in a way compatible with markdown

method = fieldnames(results{1});
measure = fieldnames(results{1}.(method{1}));
summary = {'sum','max','max'};

for i = 1:length(measure)

fprintf('\n');
fprintf('### %s \n\n', upper(measure{i}));

fprintf('| Method |');
for j = 1:length(method)
fprintf(' %8s |', method{j});
end
fprintf('\n');

fprintf('| ------ |');
for j = 1:length(method)
fprintf(' -------- |');
end
fprintf('\n');

for k = 1:size(results,1)
fprintf('| Run %2d |',k);
for j = 1:length(method)
fprintf(' %8.3f |', results{k}.(method{j}).(measure{i}));
end
fprintf('\n');
end

fprintf('| %4s   |',upper(summary{i}));
for j = 1:length(method)
fprintf(' %8.3f |', feval(summary{i},(cellfun(@(x) x.(method{j}).(measure{i}), results))));
end

fprintf('\n');

end

fprintf('\n');