Multilinear-Extension Variant Overview

• Multilinear-extension variant is the process of extending discrete or linear objects to multivariable frameworks while preserving key structures like norm attainment or transversality.
• It finds application in Banach space theory, combinatorial optimization, harmonic analysis, operator theory, and category theory using techniques such as Arens/Aron–Berner extensions and threshold-cell decompositions.
• Research in this area reveals density and approximation results along with notable obstructions, emphasizing that multilinearization is a non-formal, structurally rich enhancement of classical theories.

Across contemporary arXiv literature, the expression multilinear-extension variant denotes several non-equivalent constructions that extend a discrete function, a linear theorem, a Banach-space mapping, a Fourier extension operator, or an algebraic extension datum to a genuinely multivariable setting. The recurring pattern is the passage from an object defined on sets, vectors, operator ideals, measures, or crossed modules to a multilinear framework that preserves a specified structure such as simultaneous norm attainment, expectation-based relaxation, comonotonic cell decomposition, determinant-governed transversality, or cohomological coherence (Carando et al., 2014, Huang et al., 2021, Jost et al., 2021, Bennett et al., 2017, Wires, 2023, Aldrovandi, 2015).

Setting	Original object	Multilinear-extension variant
Banach ideals	Linear Lindenstrauss theorem	Arens/Aron–Berner extensions with simultaneous norm attainment
Combinatorial optimization	Discrete $k$-submodular or set functions	Multilinear or piecewise multilinear relaxations
Harmonic analysis	Fourier extension or multiplier bounds	Determinant-weighted multilinear identities and restriction estimates
Operator theory	Linear composition/extrapolation/maximal theorems	Multilinear mixed-norm, vector-valued, and variational extensions
Algebra and category theory	Extensions classified by cocycles	Multi-extensions, butterflies, and bimonoidal functors

1. Banach-space and bidual extension theory

In Banach-space theory, the multilinear-extension variant is centered on Arens or Aron–Berner bidual extensions of continuous multilinear mappings. For Banach spaces $E_1,\dots,E_n,F$, the space $L(E_1\times\cdots\times E_n;F)$ is equipped with the norm

$$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$

Given $T\in L(E_1\times\cdots\times E_n;F)$, there are $n!$ canonical Arens extensions $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$, indexed by permutations $k\in S_n$, defined by iterated weak$^*$ limits: $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$ The central result is a multilinear Lindenstrauss theorem inside Banach multilinear ideals: if a Banach ideal of $E_1,\dots,E_n,F$0-linear forms is stable at the underlying product space, then forms whose Arens extensions simultaneously attain their norms at a common bidual point are dense in the ideal norm (Carando et al., 2014).

The theorem has several refinements. A regular Banach ideal admits a vector-valued version through an associated ideal of $E_1,\dots,E_n,F$1-linear forms; for trilinear forms, the stability hypothesis is automatic; and for regular bilinear ideals, bilinear maps whose two Arens extensions attain their norms at the same pair are dense. The symmetric setting yields a corresponding statement for symmetric multilinear maps and homogeneous polynomials: if $E_1,\dots,E_n,F$2 is separable with the approximation property and $E_1,\dots,E_n,F$3 is either a dual space or has property (3) of Lindenstrauss, then every $E_1,\dots,E_n,F$4 can be approximated by symmetric multilinear maps whose Arens extensions simultaneously attain their norms, equivalently by polynomials whose Aron–Berner extensions attain their norms on the diagonal (Carando et al., 2014).

The structural mechanism is stability. For multilinear forms, stability is expressed through operators $E_1,\dots,E_n,F$5 built from coordinate projections $E_1,\dots,E_n,F$6 and the estimate

$E_1,\dots,E_n,F$7

This supports a perturbative iteration that preserves the ideal and forces common norm attainment of all $E_1,\dots,E_n,F$8 Arens extensions. The covered classes include minimal ideals, ideals dual to associative tensor norms, multiplicative sequences of symmetric ideals, nuclear, integral, extendible, and multiple $E_1,\dots,E_n,F$9-summing multilinear forms, as well as Hilbert–Schmidt and multilinear Schatten classes on Hilbert spaces. The paper also records a structural limitation: the multiple $L(E_1\times\cdots\times E_n;F)$0-summing ideal is multiplicative, hence stable, but is not dual to any associative tensor norm (Carando et al., 2014).

A parallel bidual-extension problem appears for almost Dunford–Pettis multilinear operators on Banach lattices. An $L(E_1\times\cdots\times E_n;F)$1-linear operator is almost Dunford–Pettis if disjoint weakly null sequences in one coordinate are sent to norm-null sequences, with the other coordinates bounded. Aron–Berner extensions preserve this property under specific lattice hypotheses: finite-rank multilinear operators keep finite rank after extension; positive multilinear operators with genuine separately weak$L(E_1\times\cdots\times E_n;F)$2-weak$L(E_1\times\cdots\times E_n;F)$3 continuous bidual extensions are separately almost Dunford–Pettis when the domain duals have the positive Schur property and the range has order continuous norm; and several bilinear and higher-order positive results are proved under order continuity or dual positive Schur assumptions (Botelho et al., 2022). The same paper shows that preservation can fail sharply: a positive separately almost Dunford–Pettis operator can have a unique Aron–Berner extension that is not separately almost Dunford–Pettis, and when the range contains $L(E_1\times\cdots\times E_n;F)$4, preservation for all positive multilinear operators is equivalent to the positive Schur property in some domain bidual (Botelho et al., 2022).

2. Continuous relaxations of discrete problems

In combinatorial optimization, the multilinear-extension variant converts a discrete objective into a continuous one on a polyhedral or simplex-like domain. For a $L(E_1\times\cdots\times E_n;F)$5-submodular function $L(E_1\times\cdots\times E_n;F)$6, the extension is defined on

$L(E_1\times\cdots\times E_n;F)$7

by independent per-item label sampling: $L(E_1\times\cdots\times E_n;F)$8 This extension is multilinear in each coordinate, has zero within-item second derivatives, has element-wise non-positive Hessian across different items, and satisfies pairwise monotonicity of partial derivatives. A key estimate is the approximate-linearity bound

$L(E_1\times\cdots\times E_n;F)$9

when $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$0. These properties support Frank–Wolfe-type algorithms for monotone and non-monotone $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$1-submodular maximization under matroid, knapsack, and mixed conjunction constraints $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$2, with a sampling-based gradient oracle and rounding procedures that reduce to submodular multilinear rounding after random label selection (Huang et al., 2021).

The resulting approximation guarantees are explicit. For monotone objectives, the framework yields $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$3-approximation under a single matroid or $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$4 knapsacks and $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$5-approximation for the intersection of $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$6 matroids and $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$7 knapsacks. For non-monotone objectives, the corresponding guarantees are $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$8 and $$\|T\|=\sup\{\|T(x_1,\ldots,x_n)\|:x_i\in B_{E_i}\}.$$9. The analysis replaces the coordinatewise-max closure used in classical submodular multilinear extension by convex-combination closure in $T\in L(E_1\times\cdots\times E_n;F)$0, and uses pairwise monotonicity to correct non-monotone directions (Huang et al., 2021).

A different discrete-to-continuous construction is the piecewise multilinear extension of set functions. For $T\in L(E_1\times\cdots\times E_n;F)$1, the extension $T\in L(E_1\times\cdots\times E_n;F)$2 is defined on $T\in L(E_1\times\cdots\times E_n;F)$3 by sorting each block, forming threshold superlevel sets $T\in L(E_1\times\cdots\times E_n;F)$4, and summing over cells: $T\in L(E_1\times\cdots\times E_n;F)$5 This extension is $T\in L(E_1\times\cdots\times E_n;F)$6-homogeneous, piecewise multilinear on comonotone cells, and reduces to the Lovász extension when one block varies and the others are frozen at characteristic vectors. The paper emphasizes a contrast with the classical multilinear extension on $T\in L(E_1\times\cdots\times E_n;F)$7: the classical version is globally multilinear, while the piecewise multilinear version is defined on all of $T\in L(E_1\times\cdots\times E_n;F)$8 and is multilinear only on each comonotone region (Jost et al., 2021).

The piecewise variant is used to prove exact discrete-to-continuous min–max equalities through the notion of perfect domain pairs, to identify inclusion chains of sets with comonotonicity of vectors, and to show that slice rank and rank are preserved under extension. The same framework leads to a spectral theory for pairs $T\in L(E_1\times\cdots\times E_n;F)$9 of locally Lipschitz $n!$0-homogeneous functions, with eigenpairs defined by

$n!$1

and to min–max characterizations, mountain-pass formulas, nodal domain inequalities, duality theorems, and a correspondence between nonzero eigenvalues of the vertex $n!$2-Laplacian and the edge $n!$3-Laplacian (Jost et al., 2021).

3. Harmonic-analytic extension operators and multiplier theorems

In harmonic analysis, the phrase refers both to multilinear Fourier extension identities and to multilinear restriction or multiplier theorems. For a hypersurface $n!$4, the Fourier extension operator is

$n!$5

For $n!$6 hypersurfaces with phases $n!$7, the determinant of the matrix whose rows are $n!$8 is the Jacobian governing a multilinear identity. The basic formula states that, under nonvanishing of this determinant on the relevant supports,

$n!$9

For the paraboloid, the determinant becomes the volume form $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$0, yielding higher-dimensional extensions of the Ozawa–Tsutsumi identity for free Schrödinger solutions. The same paper treats variable-coefficient oscillatory integrals under a multilinear transversality condition expressed through the determinant of $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$1 vectors (Bennett et al., 2017).

A different Fourier-extension variant concerns fractal measures. For compactly supported measures $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$2 on $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$3, the extension operator is

$${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$4

If the convolution $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$5 belongs to $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$6, with

$${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$7

then

$${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$8

This criterion interpolates transversality through $${}^{A_k}T:E_1^{**}\times\cdots\times E_n^{**}\to F^{**}$$9-integrability of convolution rather than smooth curvature. The paper also proves necessary conditions from upper box dimension and from Fourier decay plus a ball condition, showing that the admissible $k\in S_n$0-range becomes more restrictive when the convolution is singular (Oliveira et al., 9 Feb 2026).

A third analytic variant is a multilinear Hörmander multiplier theorem with a Lorentz–Sobolev condition. For an $k\in S_n$1-linear Fourier multiplier $k\in S_n$2, if $k\in S_n$3 and $k\in S_n$4 lies in $k\in S_n$5, then

$k\in S_n$6

The Lorentz–Sobolev space $k\in S_n$7 is optimal in the precise sense that it cannot be replaced by $k\in S_n$8 for $k\in S_n$9 or by $^*$0 for $^*$1 (Grafakos et al., 2020).

4. Extensions of operator theorems to mixed-norm, vector-valued, and noncommutative settings

Several papers use a multilinear-extension variant to transport a linear or scalar theorem to richer function-space environments. One example is the Kenig–Stein multilinear fractional integral

$^*$2

which is linearized as an operator $^*$3 on mixed-norm spaces $^*$4. The paper gives a complete characterization of boundedness $^*$5 under the homogeneity relation

$^*$6

together with relaxed structural assumptions on the block matrix $^*$7 defining the linear maps $^*$8. For $^*$9 and for $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$0, the characterization is necessary and sufficient (Chen et al., 2019).

A second strand extends Rubio de Francia extrapolation to the multilinear limited-range and vector-valued setting. If an $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$1-(sub)linear operator satisfies scalar weighted bounds

$${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$2

for weights $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$3, then, under $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$4 and $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$5, the same bounds extend to Bochner spaces: $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$6 This framework yields vector-valued bounds for the bilinear Hilbert transform, multilinear Calderón–Zygmund operators, Fourier multipliers, Bochner–Riesz operators, and spherical maximal operators (Lorist et al., 2017).

The same extension logic appears in noncommutative Calderón–Zygmund theory. Standard multilinear singular integrals are extended to tuples of UMD spaces tied by a natural product structure, including pointwise products in UMD function lattices and operator composition in Schatten–von Neumann classes. The main result gives sparse domination and $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$7 bounds for these Banach-valued extensions without assuming the Rademacher maximal function property. The resulting generality covers noncommutative $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$8 spaces and yields new fractional Leibniz rules, including Schatten-class versions (Plinio et al., 2019).

A different theorem of this type is a multilinear, multiparameter, strong $${}^{A_k}T(x_1^{**},\ldots,x_n^{**}) = w^*-\lim_{\alpha_{k(1)}}\cdots w^*-\lim_{\alpha_{k(n)}} J_F\!\big(T(x_1^{\alpha_1},\ldots,x_n^{\alpha_n})\big).$$9-variational extension of the Christ–Kiselev maximal theorem. If

$E_1,\dots,E_n,F$00

is bounded, then the truncations associated with products of filtrations satisfy strong $E_1,\dots,E_n,F$01-variation bounds whenever $E_1,\dots,E_n,F$02 and $E_1,\dots,E_n,F$03: $E_1,\dots,E_n,F$04 The paper also proves failure at the endpoint $E_1,\dots,E_n,F$05 for Fourier truncations when $E_1,\dots,E_n,F$06 (Dabhi, 28 Mar 2025).

At a more structural level, Pietsch’s composition theorem for absolutely summing linear operators has a “natural perfect extension” to multilinear and polynomial mappings via the composition multi-ideal $E_1,\dots,E_n,F$07. If $E_1,\dots,E_n,F$08, then

$E_1,\dots,E_n,F$09

with the analogous statement for homogeneous polynomial ideals. The same paper studies cotype-based refinements for concrete multilinear classes such as dominated and multiple summing operators (Bernardino et al., 2011).

5. Algebraic and categorical multi-extensions

In universal algebra, a multilinear-extension variant classifies extensions in varieties of $E_1,\dots,E_n,F$10-modules equipped with multilinear operations. For a datum $E_1,\dots,E_n,F$11, a $E_1,\dots,E_n,F$12-cocycle $E_1,\dots,E_n,F$13 produces an extension $E_1,\dots,E_n,F$14 on the set $E_1,\dots,E_n,F$15 with operations

$E_1,\dots,E_n,F$16

and multilinear operations involving action terms $E_1,\dots,E_n,F$17 and strict terms $E_1,\dots,E_n,F$18. The paper proves that $E_1,\dots,E_n,F$19-compatible $E_1,\dots,E_n,F$20-cocycles classify extensions up to equivalence through $E_1,\dots,E_n,F$21, and in the affine case obtain an additive cohomology theory $E_1,\dots,E_n,F$22. It further establishes a Wells-type exact sequence

$E_1,\dots,E_n,F$23

for group-trivial affine extensions, and a low-dimensional Hochschild–Serre exact sequence

$E_1,\dots,E_n,F$24

for an extension with an additional affine action (Wires, 2023).

In higher category theory, bimonoidal functors are encoded by nonabelian biextensions and their $E_1,\dots,E_n,F$25-variable analogues. For braided crossed modules, a biextension of $E_1,\dots,E_n,F$26 and $E_1,\dots,E_n,F$27 by $E_1,\dots,E_n,F$28 is a $E_1,\dots,E_n,F$29-torsor over $E_1,\dots,E_n,F$30 with two partial product laws and an interchange constraint governed by the braiding. A butterfly adds compatible trivializations on pullbacks, and there is an equivalence between butterflies and bimonoidal functors. This extends to $E_1,\dots,E_n,F$31-butterflies and $E_1,\dots,E_n,F$32-variable bimonoidal functors, with a composition law given by wing juxtaposition. The resulting multilinear functor calculus is organized as an equivalence of bi-multicategories, and ring-like stacks are classified by third Mac Lane cohomology $E_1,\dots,E_n,F$33 (Aldrovandi, 2015).

6. Obstructions, sharpness, and other multilinear extensions

A repeated theme is that multilinear extension does not automatically preserve the linear or scalar theory. In Banach ideals, qualitative density holds but the Lindenstrauss–Bishop–Phelps–Bollobás quantitative property fails on $E_1,\dots,E_n,F$34; in lattice theory, Aron–Berner extension can destroy the almost Dunford–Pettis property unless positive Schur or order-continuity hypotheses are imposed; in Christ–Kiselev theory, strong $E_1,\dots,E_n,F$35-variation fails at $E_1,\dots,E_n,F$36 for $E_1,\dots,E_n,F$37; in multilinear fractal restriction, box-dimension and Fourier-decay Knapp-type obstructions can force a stricter $E_1,\dots,E_n,F$38-range; and in the Lorentz–Sobolev multiplier theorem, the endpoint space $E_1,\dots,E_n,F$39 is genuinely sharp (Carando et al., 2014, Botelho et al., 2022, Dabhi, 28 Mar 2025, Oliveira et al., 9 Feb 2026, Grafakos et al., 2020). These results rule out the assumption that multilinearization is merely formal.

Another strand extends a classical mean-value theorem rather than an operator estimate. For bounded multiplicative functions $E_1,\dots,E_n,F$40 and linear forms $E_1,\dots,E_n,F$41, the average

$E_1,\dots,E_n,F$42

has an asymptotic of Halász type: $E_1,\dots,E_n,F$43 When the linear forms are pairwise independent, one can take $E_1,\dots,E_n,F$44 and $E_1,\dots,E_n,F$45, where $E_1,\dots,E_n,F$46 and $E_1,\dots,E_n,F$47 are the parameters from the one-dimensional Halász–Delange theory. The same paper proves convergence for real-valued multiplicative functions, convergence of the modulus for bounded complex-valued multiplicative functions, and convergence of conjugate-paired multilinear averages; it presents the result as a two-dimensional variant of Elliott’s conjecture (Frantzikinakis et al., 2015).

Taken together, these papers show that multilinear-extension variant is best understood as a family resemblance across fields. In each case, the extension is controlled by a rigid compatibility principle: weak$E_1,\dots,E_n,F$48 iterated limits in Banach-space theory, independent sampling or threshold-cell decomposition in optimization, determinant or convolution transversality in harmonic analysis, weighted extrapolation or product-structure duality in operator theory, and cocycle or butterfly coherence in algebra and category theory.