Tensor Product Amplification Overview

• Tensor Product Amplification is a technique that applies tensor powers or analogous polynomial maps to magnify subtle structural features in data representations.
• It converts difficult nonlinear quantities into multiplicative forms, enabling efficient extraction of low-rank signals, border-rank gaps, and spectral properties.
• The approach pairs amplification with projection or testing mechanisms to isolate enhanced features while trading off increased dimensionality and computational complexity.

Searching arXiv for the cited works to ground the article in the relevant literature.

• arXiv search query: Tensor Denoising via Amplification and Stable Rank Methods
• arXiv search query: Tensor Products and Hyperdimensional Computing
• arXiv search query: Tensor rank is not multiplicative under the tensor product
• arXiv search query: Tensor Amplification and Spectral Transfer for Sidorenko-Type Inequalities
• arXiv search query: Testing product states, quantum Merlin-Arthur games and tensor optimisation
• arXiv search query: Derandomised tensor product gap amplification for quantum Hamiltonians Tensor product amplification denotes a family of constructions in which tensor products, tensor powers, or tensor-power-like polynomial maps are used to magnify structural features that are weak or hard to access in the original representation. In recent arXiv literature the phrase appears in several technically distinct settings: low-rank tensor denoising, vector symbolic architectures and hyperdimensional computing, tensor rank and border rank in algebraic complexity, Sidorenko-type inequalities for graphons, quantum product-state testing and multi-prover verification, and gap amplification for local Hamiltonians (Gryak et al., 2023, Qiu, 2023, Christandl et al., 2017, Zhao, 2 Jul 2026, Harrow et al., 2010, Bergamaschi et al., 1 Oct 2025). A plausible unifying description is that tensoring preserves a multiplicative quantity—such as a norm, a capacity parameter, a homomorphism density, a spectral radius, or a promise-gap profile—while making low-rank structure, tuple separability, border-rank gaps, irregularity, non-productness, or energy defects more detectable.

1. General conception and recurring structure

The term does not denote a single standard operator. In some works it means repeated application of a polynomial map that is only “tensor power–like,” rather than a literal tensor power; in others it means the ordinary tensor product $t \otimes s$, the tensor Kronecker product $t \boxtimes s$, graphon tensor powers $W^{\otimes k}$, or the Hamiltonian transformation

$\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$

The common mechanism is amplification through dimensional or algebraic blow-up.

Three motifs recur. First, tensoring often converts a difficult nonlinear quantity into a more tractable object with multiplicative behavior. Examples include amplified norms approximating the tensor spectral norm, the factorization

$$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$

and the transformation of a Hamiltonian eigenvalue $\lambda$ into $1-(1-\lambda)^t$. Second, amplification is usually paired with a projection, restriction, or test that isolates the amplified feature: Frobenius projections in denoising, linear readout maps in VSA/HDC, principal restrictions in graphon arguments, swap-test acceptance in product-state testing, and expander-subsampled path clauses in Hamiltonian amplification. Third, amplification almost always trades a stronger signal for larger ambient complexity: higher tensor order, larger dimension, more local terms, or increased locality.

A recurring misconception is to identify tensor product amplification with naive tensor powers alone. Several of the cited works explicitly reject that identification. The denoising framework uses polynomial maps built from outer products, contractions, permutations, and sums rather than a straightforward $\mathcal T^{\otimes k}$; the VSA/HDC results show that compressed bindings are linear images of tensor products rather than substitutes of equal expressive power; and the algebraic-complexity literature sharply distinguishes the order-increasing tensor product from the order-preserving tensor Kronecker product.

2. Polynomial tensor amplification in low-rank tensor denoising

In "Tensor Denoising via Amplification and Stable Rank Methods" (Gryak et al., 2023), tensor amplification is introduced as a tensor-native analogue of spectral power amplification. The guiding matrix map is

$\phi:A\mapsto AA^\top A,$

whose singular values are $\lambda_1^3,\dots,\lambda_r^3$ when $t \boxtimes s$0 has singular values $t \boxtimes s$1. Repeated application accentuates large singular values relative to small ones. The tensor version replaces $t \boxtimes s$2 by degree-$t \boxtimes s$3 tensor polynomials $t \boxtimes s$4 built from outer products, contractions, permutations, and sums. These maps are polynomial in the tensor entries, amplify low-rank structure, and define norms $t \boxtimes s$5 satisfying

$t \boxtimes s$6

The paper uses two explicit third-order amplification maps from Tokcan–Derksen (2021): $t \boxtimes s$7, a degree-4 amplification map, and $t \boxtimes s$8, which empirically gives a better approximation to the spectral norm than $t \boxtimes s$9. For fourth-order tensors it uses a compatible version of $W^{\otimes k}$0, while an analogue of $W^{\otimes k}$1 is not yet available for order $W^{\otimes k}$2. The amplified iterate is written

$W^{\otimes k}$3

The denoising problem is posed as $W^{\otimes k}$4, with $W^{\otimes k}$5 low rank and $W^{\otimes k}$6 noise, inside a dual-norm framework influenced by Derksen’s analysis of Pareto-efficient decompositions. Algorithm 1 computes $W^{\otimes k}$7, normalizes $W^{\otimes k}$8, projects the residual away from $W^{\otimes k}$9, and outputs $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$0. This is explicitly described as akin to a tensor power method, but with algebraic amplification instead of direct optimization over rank-1 factors. The practical significance is twofold. First, the method keeps computation in tensor space, rather than flattening. Second, it approximates NP-hard tensor spectral and nuclear norms through tractable tensor operations.

The same paper introduces stable slice rank and stable $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$1-rank as robust rank surrogates that are conceptually parallel to amplification, though not themselves amplification maps. Stable slice rank is defined by

$\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$2

and the operational stable $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$3-rank reported by Algorithm 3 is

$\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$4

Empirically, the amplification-based method is often best or close to best for rank-1 and very low-rank synthetic tensors, especially at low SNR, while the stable $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$5-rank method is the only one that yields consistent denoising gains on the ECG-derived tensors. This suggests that, in this setting, tensor amplification is most naturally a rank-1 or small-$\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$6 spectral extraction procedure rather than a general high-rank tensor approximation method.

3. Tensor product representations in VSA and hyperdimensional computing

In "Tensor Products and Hyperdimensional Computing" (Qiu, 2023), tensor products are treated as the central representation for binding in vector symbolic architectures and hyperdimensional computing. The point of departure is the superposition principle: sets are represented by linear sums,

$\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$7

If a binding operation $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$8 respects superposition in each argument, then $\mathrm{TP}_t(H)=I-(I-H)^{\otimes t}.$9 is multilinear. By the universal property of the tensor product, this forces a factorization through the canonical map

$$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$0

so there exists a unique linear map $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$1 such that

$$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$2

This establishes tensor product binding as the most general superposition-compatible binding. It also clarifies why the dimensional blow-up $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$3 is not treated as an incidental cost. If $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$4 is $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$5-dimensional, then $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$6 is $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$7-dimensional, and an orthonormal basis in $$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$8 yields an orthonormal basis of tuple representations

$$t(H,W^{\otimes k})=t(H,W)^k,\qquad \rho(W^{\otimes k})=\rho(W)^k,$$9

The inner product factorizes,

$\lambda$0

so any component mismatch makes the composite structures orthogonal.

The paper proves stronger minimality statements. Up to linear isomorphism, $\lambda$1 is the unique second-order representation of minimal dimension that admits left and right unbinding operations which respect superposition, unbind correctly, and have zero error for mismatched pairs. More generally, for $\lambda$2 orthonormal embeddings, accurate detection of all $\lambda$3-tuples requires dimension at least $\lambda$4, and $\lambda$5 is, up to isomorphism, the unique $\lambda$6-order representation of minimal dimension that achieves this bound.

This yields an explicit amplification of representational capacity and decodability. For normalized Rademacher embeddings, the paper compares Hadamard binding in dimension $\lambda$7 against tensor binding in dimension $\lambda$8. In the Hadamard case,

$\lambda$9

so capacity scales like $1-(1-\lambda)^t$0. In the tensor case,

$1-(1-\lambda)^t$1

so capacity scales like $1-(1-\lambda)^t$2. The paper situates these results relative to Smolensky’s Tensor Product Representations and Plate’s Holographic Reduced Representations: compressed bindings such as Hadamard products, circular convolution, and XOR are linear images of tensor products, but they do not preserve the full errorless unbinding and detection properties.

4. Rank, border rank, and asymptotic amplification under tensor powers

In algebraic complexity, "Tensor rank is not multiplicative under the tensor product" (Christandl et al., 2017) analyzes tensor product amplification as an asymptotic mechanism acting on rank gaps. For tensors $1-(1-\lambda)^t$3 and $1-(1-\lambda)^t$4,

$1-(1-\lambda)^t$5

so tensor rank is submultiplicative under the order-increasing tensor product. The main result is that it is not multiplicative in general. If a tensor $1-(1-\lambda)^t$6 has border rank strictly smaller than rank,

$1-(1-\lambda)^t$7

then for sufficiently large $1-(1-\lambda)^t$8,

$1-(1-\lambda)^t$9

The amplification mechanism proceeds through degenerations. If $\mathcal T^{\otimes k}$0 denotes a refined border-rank quantity with bounded error degree $\mathcal T^{\otimes k}$1, then the paper proves

$\mathcal T^{\otimes k}$2

A small one-copy gap between $\mathcal T^{\otimes k}$3 and $\mathcal T^{\otimes k}$4 is therefore magnified by tensor powers into strict submultiplicativity of exact rank. The asymptotic consequence is that

$\mathcal T^{\otimes k}$5

and the exponential growth rate of rank under tensor powers is controlled by border rank rather than one-shot rank.

The paper gives explicit examples. For the $\mathcal T^{\otimes k}$6 tensors, $\mathcal T^{\otimes k}$7 while $\mathcal T^{\otimes k}$8, which yields

$\mathcal T^{\otimes k}$9

For $\phi:A\mapsto AA^\top A,$0, over fields with $\phi:A\mapsto AA^\top A,$1 and $\phi:A\mapsto AA^\top A,$2,

$\phi:A\mapsto AA^\top A,$3

The same pattern appears for Strassen tensors $\phi:A\mapsto AA^\top A,$4 and the matrix multiplication tensor $\phi:A\mapsto AA^\top A,$5.

A central conceptual distinction in this literature is the difference between the order-increasing tensor product $\phi:A\mapsto AA^\top A,$6 and the order-preserving tensor Kronecker product $\phi:A\mapsto AA^\top A,$7. Nonmultiplicativity for $\phi:A\mapsto AA^\top A,$8 was already known from Strassen’s work, whereas the novelty here is nonmultiplicativity for $\phi:A\mapsto AA^\top A,$9. Another significant contrast is methodological: lower bounds on border rank obtained from generalised flattenings, including Young flattenings, multiply under tensor powers even though actual rank need not. This creates a systematic tension between multiplicative lower-bound methods and the true behavior of rank.

5. Graphon tensor amplification for Sidorenko-type inequalities

In "Tensor Amplification and Spectral Transfer for Sidorenko-Type Inequalities" (Zhao, 2 Jul 2026), tensor amplification is formulated for graphon classes closed under tensor powers and normalized principal restrictions. For a graphon $\lambda_1^3,\dots,\lambda_r^3$0, the $\lambda_1^3,\dots,\lambda_r^3$1-fold tensor power is

$\lambda_1^3,\dots,\lambda_r^3$2

and satisfies

$\lambda_1^3,\dots,\lambda_r^3$3

The class $\lambda_1^3,\dots,\lambda_r^3$4 is called admissible if it is closed under these tensor powers and under normalized principal restrictions $\lambda_1^3,\dots,\lambda_r^3$5.

The framework isolates two amplification mechanisms. Degree-biased tensor amplification uses the degree-biased measure

$\lambda_1^3,\dots,\lambda_r^3$6

If $\lambda_1^3,\dots,\lambda_r^3$7 is not $\lambda_1^3,\dots,\lambda_r^3$8-regular, then there are sets $\lambda_1^3,\dots,\lambda_r^3$9 with

$t \boxtimes s$00

Thus tensor powers do not improve the normalized Sidorenko ratio directly, but they make degree irregularity concentrate on very small principal restrictions. Applying the $t \boxtimes s$01-Sidorenko inequality to those restrictions yields the equality-case regularization theorem: if $t \boxtimes s$02 is a non-matching $t \boxtimes s$03-Sidorenko graph and

$t \boxtimes s$04

then $t \boxtimes s$05 is $t \boxtimes s$06-regular. Consequently, relative forcing is equivalent to relative regular-forcing for every non-matching $t \boxtimes s$07-Sidorenko graph.

Perron-biased tensor amplification detects spectral structure. For

$t \boxtimes s$08

the weak spectral regularization theorem proves

$t \boxtimes s$09

This yields the spectral transfer principle: for admissible $t \boxtimes s$10 and graphs with $t \boxtimes s$11, ordinary $t \boxtimes s$12-Sidorenko is equivalent to the spectral inequality

$t \boxtimes s$13

for every non-zero $t \boxtimes s$14.

The paper applies the framework to doubly nonnegative graphons and bounded doubly nonnegative kernels, showing that the DNN class is admissible. This produces spectral equivalences for Sidorenko-good graphs in the range $t \boxtimes s$15 and identifies Sidorenko-good forcing with regular-KNRS forcing for non-matching Sidorenko-good graphs. Here tensor amplification is neither a rank surrogate nor a representation theorem; it is an error-amplification method that combines multiplicative tensor powers with carefully chosen principal restrictions.

6. Product-state testing, tensor optimisation, and QMA amplification

In quantum information and quantum complexity, "Testing product states, quantum Merlin-Arthur games and tensor optimisation" (Harrow et al., 2010) uses tensor-product structure as the property being amplified and certified. For a pure state on

$t \boxtimes s$16

the product-state test takes two copies of the state and runs a swap test on each corresponding subsystem pair, accepting iff all swap tests accept. For a mixed state $t \boxtimes s$17, the acceptance probability is

$t \boxtimes s$18

If

$t \boxtimes s$19

is the maximum squared overlap with a product state, then

$t \boxtimes s$20

Accordingly, $t \boxtimes s$21 for small $t \boxtimes s$22, with no dependence on $t \boxtimes s$23 or on local dimensions.

The analysis is embedded in a stability theorem for the depolarising channel. The product test acceptance is proportional to the output purity of $t \boxtimes s$24, so near-maximal output purity forces approximate product structure. This turns non-productness into a dimension-independent constant drop in an efficiently measurable observable.

The same test drives complexity-theoretic amplification. It is used to simulate $t \boxtimes s$25 by $t \boxtimes s$26 with

$t \boxtimes s$27

and, together with separable parallel repetition,

$t \boxtimes s$28

Combining these ingredients yields $t \boxtimes s$29 for all $t \boxtimes s$30. The paper also connects tensor-product optimisation to $t \boxtimes s$31 via the support function $t \boxtimes s$32 and the injective tensor norm

$t \boxtimes s$33

Here amplification is not primarily about taking tensor powers of an input tensor; it is about converting a bounded geometric gap from the set of product states into a constant algorithmic and complexity-theoretic gap.

7. Derandomised tensor product gap amplification for Hamiltonians

"Derandomised tensor product gap amplification for quantum Hamiltonians" (Bergamaschi et al., 1 Oct 2025) returns to literal tensor products. For a normalized local Hamiltonian $t \boxtimes s$34, the basic transformation is

$t \boxtimes s$35

If $t \boxtimes s$36 has eigenvalues $t \boxtimes s$37, then $t \boxtimes s$38 has eigenvalues $t \boxtimes s$39. For small $t \boxtimes s$40, this is approximately $t \boxtimes s$41, so the ground-energy gap is amplified roughly by a factor $t \boxtimes s$42. The obstacle is combinatorial blow-up: if

$t \boxtimes s$43

then the naive expansion has $t \boxtimes s$44 terms.

The paper derandomises this tensor product amplification by using random walks on expander graphs. For a layered Hamiltonian

$t \boxtimes s$45

each layer is equipped with a regular spectral expander $t \boxtimes s$46. Paths $t \boxtimes s$47 of length $t \boxtimes s$48 in $t \boxtimes s$49 define amplified clauses

$t \boxtimes s$50

and the derandomised amplified Hamiltonian is

$t \boxtimes s$51

This reduces the number of terms from $t \boxtimes s$52 to $t \boxtimes s$53, where $t \boxtimes s$54 is the expander degree.

The main amplification theorem concerns a derandomised $t \boxtimes s$55-fold amplification $t \boxtimes s$56. It is $t \boxtimes s$57-local, has $t \boxtimes s$58 terms, satisfies the completeness bound

$t \boxtimes s$59

and the soundness bound

$t \boxtimes s$60

The soundness analysis uses a new de Finetti–inspired technique together with expander mixing and an auxiliary energy measurement. The result is iterability: starting from QMA-hard layered local Hamiltonians with inverse-polynomial promise gap, the paper constructs QMA-hard families with constant promise gap and higher locality. It explicitly does not solve the constant-locality quantum PCP conjecture; locality increases with amplification, and the paper frames the outcome instead as a locality-gap tradeoff and a step toward a quantum analogue of Dinur-style gap amplification.

Across these settings, tensor product amplification is best understood not as a single theorem but as a reusable paradigm. It enlarges representation space or instance size in a controlled algebraic way, preserves a multiplicative backbone, and then exploits the amplified object to expose structure that is inaccessible or unstable at the original scale. The specific object being amplified—spectral mass, tuple separability, border-rank gaps, degree bias, non-productness, or energy defects—depends on the domain, but the recurring strategy is strikingly consistent.