1 of 36

The Two Parallel Worlds of Algorithms for Tensor Decomposition �

Pravesh Kothari

Princeton

WACT 2026

based on joint work with Alex Wein (UC Davis) and Ankur Moitra (MIT)

2 of 36

Tensor Rank/Decomposition

 

 

 

*of order 3. We’ll focus only on order 3.

Natural generalization of matrix (i.e., tensors of order 2) rank.

Question: Efficient algorithms for computing rank r of T?

…for decomposing into a sum of r rank 1 tensors?

…“canonical polyadic” or CP rank

Captures the circuit complexity of bilinear computation.

[Strassen’60s]

3 of 36

Tensor Rank: Algorithms?

Question: Efficient algorithms for computing rank/rank decomposition of T?

Why?

Learn parameters of a statistical model by estimating its (low-degree) moment tensors from samples.

e.g., Mixture of spherical Gaussians

[Hsu,Kakade’11]

 

 

 

can be computed easily from samples

Method of Moments

[Pearson’1890s]

apply tensor decomposition!

4 of 36

Question: Efficient algorithms for computing rank/rank decomposition of T?

e.g., Mixture of spherical Gaussians

[Hsu,Kakade’11]

 

 

 

Are we done?

Tensor Rank: Algorithms?

5 of 36

Question: Efficient algorithms for computing rank/rank decomposition of T?

e.g., Mixture of spherical Gaussians

[Hsu,Kakade’11]

 

 

 

Tensor Rank: Uniqueness

Rank decomposition may not be unique!

 

Fact: Tensor rank decompositions are “generically” unique.

6 of 36

Tensor Rank: Algorithms?

Matrix rank is polytime computable (Gaussian elimination).

Question: Efficient algorithms for computing rank/rank decomposition of T?

Tensor rank is NP-hard.

[Håstad’90,…]

Hope: hardness only holds for tensors with “pathological” components.

7 of 36

Tensor Rank: Jennrich’s Algorithm

 

8 of 36

Tensor Rank: Jennrich’s Algorithm

 

9 of 36

Tensor Rank: Jennrich’s Algorithm

 

 

 

 

Obs2: If g,h are random Gaussians then all eigs are distinct.

 

 

 

Note: The eigenvectors are not pairwise orthogonal, in general.

10 of 36

Tensor Rank: Uniqueness

 

 

the min rank decomposition is unique up to scaling + permutation.

Further,

[Harshmann’70]

All known algorithmic results for tensor decomposition also establish uniqueness!

11 of 36

Tensor Rank: Uniqueness

 

[Harshmann’70]

The failure case is when the components are a root of a non-zero polynomial.

Works for almost all inputs. Enough to beat NP-hardness!

“generic components”

 

overcomplete regime

 

12 of 36

Tensor Rank: Algorithms vs Explicit Constructions

 

 

explicit constructions* imply formula size lower bounds.

[Raz’10]

 

 

the min rank decomposition is unique up to scaling + permutation.

Further,

[Harshmann’70]

 

 

[Alexeev-Fobes-Tsimerman’11]

13 of 36

Overcomplete tensor decomposition

 

[Chen-Rademacher’20]

What we know:

[Koiran’24]

 

[K-Moitra-Wein’24]

 

“commuting extensions”

“Koszul-Young Flattenings”

 

 

completely elementary analysis.

[Kruskal’77]

 

can finally match/go beyond the Kruskal uniqueness condition

14 of 36

The main idea

15 of 36

Rank Detection

 

 

Such certificates are natural in proving, lower bounds for, say, bilinear computation.

[Garg-Kayal-Saha’20] use ideas from lower bound proofs to give reconstruction algorithms

It seems like any decomposition algorithm also naturally yields such certificates*.

Idea: “from certificates to algorithms”

“shifted partial derivative” certificates 🡪 reconstruction algorithms

[Barak,Brandao,Harrow,Kelner,Steurer,Zhou’16]

[Fleming-K-Pitassi’19]

Reminiscent of “sos certificates to algorithms” paradigm:

*We will come back to this later.

16 of 36

Rank Detection

 

 

Idea: reduce to matrix rank (reminiscent of Rohit’s talk yesterday!)

(Rank Detector Gadget)

 

 

 

17 of 36

Rank Detection

[Persu’18]

 

(Rank Detector Gadget)

 

 

 

[K-Moitra-Wein’24]

 

+ a decomposition algo based on finding rank 1 matrices planted in generic subspaces

Let’s think about constructing such gadgets.

[Strassen’83]

 

18 of 36

Trivial Flattening

 

 

 

 

 

 

 

 

 

Our goal is to beat this by a factor ~2.

19 of 36

Koszul-Young Flattening

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

For rank 1 T:

[Landsberg-Ottaviani’13]

 

 

 

 

20 of 36

Koszul-Young Flattening

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

For rank 1 T:

 

 

 

 

[Landsberg-Ottaviani’13]

 

  • Rows/Columns of A(u) are indexed by sets of size p,p+1 respectively.

 

 

21 of 36

Koszul-Young Flattening

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

For rank 1 T:

 

 

 

 

[Landsberg-Ottaviani’13]

 

 

 

 

 

Both proofs are by elementary are linear algebra.

22 of 36

Koszul-Young Flattening

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

[Landsberg-Ottaviani’13]

 

 

 

 

 

 

23 of 36

Decomposition Algorithm

 

Input:

Goal:

 

Obs:

 

Key:

We are looking for vectors that have a “rank 1” structure (when viewed as matrices)

Fact:

 

[Johnston-Lovitz-Vijayraghavan’22]

 

 

[Barak-K-Steurer’17]

Closely related to the quantum “best separable state” problem.

[Aaronson-Impagliazzo-Moshkovitz’11]

24 of 36

Decomposition Algorithm

 

Input:

Goal:

 

Obs:

 

Key:

We are looking for vectors that have a “rank 1” structure (when viewed as matrices)

Fact:

 

[Johnston-Lovitz-Vijayraghavan’22]

 

 

[Barak-K-Steurer’17]

Closely related to the quantum “best separable state” problem.

[Aaronson-Impagliazzo-Moshkovitz’11]

[Dastidar-Weicht-Wein’25]

25 of 36

Overcomplete tensor decomposition

[K-Moitra-Wein’24]

 

“Koszul-Young Flattenings”

 

 

completely elementary analysis.

can finally match/go beyond the Kruskal uniqueness condition

Lower Bounds

 

 

[K-Moitra-Wein’24]

[Efremenko-Garg-Oliveira-Wigderson’18]

 

[K-Moitra-Wein’24]

26 of 36

Overcomplete tensor decomposition

[K-Moitra-Wein’24]

 

“Koszul-Young Flattenings”

 

 

completely elementary analysis.

can finally match/go beyond the Kruskal uniqueness condition

Lower Bounds

 

 

[K-Moitra-Wein’24]

[Efremenko-Garg-Oliveira-Wigderson’18]

 

[K-Moitra-Wein’24]

Beating 3n will likely allow improving the construction of .

[Alexeev-Forbers-Tsimerman’11]

27 of 36

Overcomplete tensor decomposition

 

Can stronger assumptions on the components help?

28 of 36

Tensors with Random Components

Theorem

[Ma-Shi-Steurer’15]

 

 

  • Goes far beyond the regime we could handle for generic components.
  • Does not yield an exact recovery guarantee.

How do they go (so far) beyond the results we could prove in the generic setting?

29 of 36

Tensors with Random Components

Theorem

[Ma-Shi-Steurer’15]

 

 

Key Idea:

 

Observation:

 

This is (very) not true for order 2 tensors (matrices).

Natural algorithm:

 

NP-hard in general. But instead, take SDP (“sum of squares”) relaxation.

Analysis reduces to giving a “low degree sum of squares proof” of observation.

30 of 36

Tensors with Random Components

Theorem

[Ma-Shi-Steurer’15]

 

 

Key Idea:

 

Observation:

 

This is (very) not true for order 2 tensors (matrices).

How far can we go beyond “random”?

Such an algorithm does not certify rank (so no “tensor rank” obstruction).

31 of 36

Smoothed Tensor Decomposition?

Theorem

[Ma-Shi-Steurer’15]

 

 

Question: Overcomplete tensor decomposition with smoothed components?

 

 

32 of 36

Theorem

[Ma-Shi-Steurer’15]

 

 

Key Idea:

 

Observation:

 

This is (very) not true for order 2 tensors (matrices).

The proof of the observation relies strongly on approximate orthogonality of random vectors (that is not a generic property). Does not extend to smoothed setting.

Smoothed Tensor Decomposition?

33 of 36

Theorem

[Ma-Shi-Steurer’15]

 

 

Smoothed Tensor Decomposition?

Any such result will likely need to obtain exact as opposed to approx recovery.

[Kivva-Potechin’21]

Exact recovery with same parameters for asymmetric, random components.

I’ll briefly describe their approach since it seems sensible for smoothed setting too.

34 of 36

The Kivva-Potechin Approach

Def (Injective and Nuclear Norms of Tensors)

 

 

 

Observe: This does NOT certify that the natural decomposition is a min-rank decomposition.

It only certifies that the natural decomposition is the min nuclear norm decomposition.

35 of 36

Question/Conjecture

Def (Injective and Nuclear Norms of Tensors)

 

 

36 of 36

Summary/Open Qs

 

Via “rank detection gadgets”

Are there better rank detection gadgets/flattenings?

Can we beat the tensor rank barrier for tensors with smoothed components?