Canoni
cal Polyadi
c De
compo
sition (CPD) of a third-order ten
sor i
s a minimal de
compo
sition into a
sum of rank-1 ten
sor
s. We find new mild determini
sti
c condition
s for the uniquene
ss of individual rank-1 ten
sor
s in CPD and pre
sent an algorithm to re
cover them. We
call the algorithm “algebrai
c” be
cau
se it relie
s only on
standard linear algebra. It doe
s not involve more advan
ced pro
cedure
s than the
computation of the null
spa
ce of a matrix and eigen/
singular value de
compo
sition. Simulation
s indi
cate that the new
condition
s for uniquene
ss and the working a
ssumption
s for the algorithm hold for a randomly generated <
span id="mml
si1"
cla
ss="mathml
sr
c"><
span
cla
ss="formulatext
stixSupport mathImg" data-mathURL="/
scien
ce?_ob=MathURL&_method=retrieve&_eid=1-
s2.0-S00
2437951630492X&_mathId=
si1.gif&_u
ser=111111111&_pii=S00
2437951630492X&_rdo
c=1&_i
ssn=00
243795&md5=f6de898797ba556d5fd0597b31bf6257" title="Cli
ck to view the MathML
sour
ce">I&time
s;J&time
s;K
span><
span
cla
ss="mathContainer hidden"><
span
cla
ss="mathCode">
span>
span>
span> ten
sor of rank <
span id="mml
si2"
cla
ss="mathml
sr
c"><
span
cla
ss="formulatext
stixSupport mathImg" data-mathURL="/
scien
ce?_ob=MathURL&_method=retrieve&_eid=1-
s2.0-S00
2437951630492X&_mathId=
si2.gif&_u
ser=111111111&_pii=S00
2437951630492X&_rdo
c=1&_i
ssn=00
243795&md5=4d7
c74b65baff1
cdf1ad2719b01f
c739" title="Cli
ck to view the MathML
sour
ce">R≥K≥J≥I≥2
span><
span
cla
ss="mathContainer hidden"><
span
cla
ss="mathCode">
span>
span>
span> if
R i
s bounded a
s <
span id="mml
si3"
cla
ss="mathml
sr
c">
source" class="mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si3.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=46382bf2f5161f01eaf99696e8e05785">class="imgLazyJSB inlineImage" height="20" width="365" alt="View the MathML source" title="View the MathML source" src="/sd/grey_pxl.gif" data-inlimgeid="1-s2.0-S002437951630492X-si3.gif">script>style="vertical-align:bottom" width="365" alt="View the MathML source" title="View the MathML source" src="http://origin-ars.els-cdn.com/content/image/1-s2.0-S002437951630492X-si3.gif">script><
span
cla
ss="mathContainer hidden"><
span
cla
ss="mathCode">
span>
span>
span> at lea
st for the dimen
sion
s that we have te
sted. Thi
s improve
s upon the famou
s Kru
skal bound for uniquene
ss <
span id="mml
si4"
cla
ss="mathml
sr
c"><
span
cla
ss="formulatext
stixSupport mathImg" data-mathURL="/
scien
ce?_ob=MathURL&_method=retrieve&_eid=1-
s2.0-S00
2437951630492X&_mathId=
si4.gif&_u
ser=111111111&_pii=S00
2437951630492X&_rdo
c=1&_i
ssn=00
243795&md5=
c007a50703b1af83fb3b936dd195b12e" title="Cli
ck to view the MathML
sour
ce">R≤(I+J+K&minu
s;2)/2
span><
span
cla
ss="mathContainer hidden"><
span
cla
ss="mathCode">
span>
span>
span> a
s soon a
s <
span id="mml
si5"
cla
ss="mathml
sr
c"><
span
cla
ss="formulatext
stixSupport mathImg" data-mathURL="/
scien
ce?_ob=MathURL&_method=retrieve&_eid=1-
s2.0-S00
2437951630492X&_mathId=
si5.gif&_u
ser=111111111&_pii=S00
2437951630492X&_rdo
c=1&_i
ssn=00
243795&md5=1
c26b74db1d2f003d9a97fa282740de1" title="Cli
ck to view the MathML
sour
ce">I≥3
span><
span
cla
ss="mathContainer hidden"><
span
cla
ss="mathCode">
span>
span>
span>.
sp0060">In the particular case <span id="mmlsi6" class="mathmlsrc"><span class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si6.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=ce7737c9ed7edde190c7645e4a21d2ec" title="Click to view the MathML source">R=Kspan><span class="mathContainer hidden"><span class="mathCode">span>span>span>, the new bound above is equivalent to the bound <span id="mmlsi194" class="mathmlsrc"><span class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si194.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=d43df3f83f3ee7be3a942149e8bcf315" title="Click to view the MathML source">R≤(I−1)(J−1)span><span class="mathContainer hidden"><span class="mathCode">span>span>span> which is known to be necessary and sufficient for the generic uniqueness of the CPD. An existing algebraic algorithm (based on simultaneous diagonalization of a set of matrices) computes the CPD under the more restrictive constraint <span id="mmlsi8" class="mathmlsrc"><span class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si8.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=fc633c849a1f3849b8e7eb69ba9e9b03" title="Click to view the MathML source">R(R−1)≤I(I−1)J(J−1)/2span><span class="mathContainer hidden"><span class="mathCode">span>span>span> (implying that <span id="mmlsi9" class="mathmlsrc">source" class="mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si9.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=2caf5463fa1b9b944ab5dd1f0871ed50">class="imgLazyJSB inlineImage" height="19" width="196" alt="View the MathML source" title="View the MathML source" src="/sd/grey_pxl.gif" data-inlimgeid="1-s2.0-S002437951630492X-si9.gif">script>style="vertical-align:bottom" width="196" alt="View the MathML source" title="View the MathML source" src="http://origin-ars.els-cdn.com/content/image/1-s2.0-S002437951630492X-si9.gif">script><span class="mathContainer hidden"><span class="mathCode">span>span>span>). We give an example of a low-dimensional but high-rank CPD that cannot be found by optimization-based algorithms in a reasonable amount of time while our approach takes less than a second. We demonstrate that, at least for <span id="mmlsi10" class="mathmlsrc"><span class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si10.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=486125734e02de25e433221b8d83af6f" title="Click to view the MathML source">R≤24span><span class="mathContainer hidden"><span class="mathCode">span>span>span>, our algorithm can recover the rank-1 tensors in the CPD up to <span id="mmlsi194" class="mathmlsrc"><span class="formulatext stixSupport mathImg" data-mathURL="/science?_ob=MathURL&_method=retrieve&_eid=1-s2.0-S002437951630492X&_mathId=si194.gif&_user=111111111&_pii=S002437951630492X&_rdoc=1&_issn=00243795&md5=d43df3f83f3ee7be3a942149e8bcf315" title="Click to view the MathML source">R≤(I−1)(J−1)span><span class="mathContainer hidden"><span class="mathCode">span>span>span>.