---
title: 'Krivine Schemes: Banach Spaces, Rounding, and Optimization'
url: https://www.emergentmind.com/topics/krivine-schemes
type: topic
---

# Krivine Schemes: Banach Spaces, Rounding, and Optimization

Krivine schemes are a family of mathematical and computational constructions associated with several distinct uses of Krivine’s name. In the theory of Banach spaces, a Krivine scheme records finite-dimensional normed spaces that occur with arbitrarily small distortion inside a Banach space. In Grothendieck-type inequalities and semidefinite rounding, a Krivine scheme is a randomized transformation of Euclidean vectors into signs, usually through Gaussian projections and Boolean functions, designed to preserve inner products up to a controlled factor. In polynomial optimization, Krivine–Stengle schemes form hierarchies of linear-programming positivity certificates. Related but formally different objects include Krivine Abstract Machines, Alternating Parity Krivine Automata, Krivine realizability algebras, and Krivine diffusions.

## 1. Terminology and mathematical lineages

The expression *Krivine scheme* is not associated with a single formalism. Its meaning depends on the surrounding field.

In Banach-space theory, the term refers to approximate finite-dimensional representability. A Banach space \(X\) contains \(\ell_p^n\) with distortion at most \(1+\varepsilon\) if there is an \(n\)-dimensional subspace of \(X\) linearly isomorphic to \(\ell_p^n\) with distortion at most \(1+\varepsilon\). The values of \(p\) for which this holds uniformly in \(n\) form the Krivine set, or Krivine scheme, of \(X\). This notion concerns approximate linear copies, not exact finite metric copies.

In Grothendieck theory, a Krivine scheme is a rounding procedure converting vector solutions of a semidefinite relaxation into scalar signs. Its analysis is governed by a Gaussian correlation function and, typically, by tensor-power transformations that invert this correlation function. The objective is to approximate the real Grothendieck constant \(K_G\), the least universal constant satisfying
\[
\operatorname{SDP}(A)\le K_G\operatorname{IP}(A).
\]

In polynomial optimization, a Krivine scheme is a finite-order hierarchy of Krivine–Stengle representations. A positive polynomial is represented as a nonnegative linear combination of products of normalized constraint polynomials \(g_i\) and their complements \(1-g_i\). Coefficient matching yields a linear program.

The term is also used in computational semantics. A Krivine Abstract Machine is an evaluator for lambda terms based on closures, environments, continuations, and stacks. Alternating Parity Krivine Automata combine such higher-order evaluation with alternating parity tree automata. Krivine realizability extends the machine-level apparatus into a classical realizability semantics for set theory. These formalisms are conceptually related through higher-order computation but are not instances of the Banach-space or Grothendieck-rounding meanings.

## 2. Krivine schemes in Banach-space theory

Let \(X\) and \(Y\) be Banach spaces. A finite-dimensional space \(E\) is finitely representable in \(X\) if, for every \(\varepsilon>0\), there is a linear embedding \(T:E\to X\) with
\[
\|T\|\,\|T^{-1}\|\le 1+\varepsilon.
\]
A Banach space \(Y\) is finitely representable in \(X\) when every finite-dimensional subspace of \(Y\) is finitely representable in \(X\).

The classical Krivine theorem states that if \(X\cong\ell_p\), then, for every \(n\) and \(\varepsilon>0\), \(X\) contains an \(n\)-dimensional subspace \((1+\varepsilon)\)-isomorphic to \(\ell_p^n\). Equivalently, \(\ell_p^n\) occurs in \(X\) with arbitrarily small linear distortion, uniformly in \(n\).

This approximate assertion is substantially weaker than exact finite isometric representability. Linear isomorphism preserves norms only up to multiplicative constants, whereas an isometric embedding of a finite metric subset preserves every distance exactly. The distinction is expressed by the implications
\[
X\cong\ell_p
\quad\not\Rightarrow\quad
\text{every finite subset of }\ell_p\text{ embeds isometrically into }X.
\]

For \(1<p<2\), there exists a space \(X\cong\ell_p\) into which
\[
U=\{e_1,e_2,-e_1,-e_2,0\}
\]
does not embed isometrically. The construction uses the Orlicz sequence space \(\ell_M\) with
\[
M(t)=t^p+t^r,\qquad p<r<2.
\]
Strict convexity forces the images of \(e_1\) and \(-e_1\), and of \(e_2\) and \(-e_2\), to be antipodal. The prescribed distances then imply
\[
\|x+y\|_M=\|x-y\|_M=2^{1/p},
\]
which contradicts the corresponding Clarkson inequality after separating the \(p\)- and \(r\)-power contributions.

For \(2<p<\infty\), the construction uses a modular sequence space with
\[
M_i(t)=t^p+t^{p_i},\qquad 2<p_i<p,\qquad p_i\to p
\]
sufficiently rapidly. This space is isomorphic to \(\ell_p\), but the finite set
\[
V=\left\{\pm 2^{-1/p}(e_1+e_2),\ \pm 2^{-1/p}(e_1-e_2),\ 0\right\}
\]
does not embed isometrically into it. The obstruction again follows from strict convexity and Clarkson inequalities, now in the regime \(p>2\) and \(p_i>2\) [1708.01570].

The geometric mechanism can be described using the James constant
\[
J(X)=\sup\{\min\{\|x+y\|,\|x-y\|\}:x,y\in S_X\}.
\]
The constructed spaces \(X\cong\ell_p\) have the same James constant as \(\ell_p\), but the supremum is not attained. Thus exact extremal configurations may disappear while the optimal numerical constant remains unchanged. The case \(p=2\) is exceptional: the strict coefficient inequalities used in both contradictions disappear, and the corresponding general finite-isometric question remains open in the cited discussion.

## 3. Krivine rounding and Grothendieck’s inequality

For a real matrix \(A=(a_{ij})\), define
\[
\operatorname{SDP}(A)
=
\sup_{\|x_i\|=\|y_j\|=1}
\sum_{i,j}a_{ij}\langle x_i,y_j\rangle
\]
and
\[
\operatorname{IP}(A)
=
\max_{\varepsilon_i,\delta_j\in\{-1,1\}}
\sum_{i,j}a_{ij}\varepsilon_i\delta_j.
\]
Grothendieck’s inequality asserts that
\[
\operatorname{SDP}(A)\le K_G\operatorname{IP}(A),
\]
where \(K_G\) is the real Grothendieck constant.

A \(k\)-dimensional Krivine scheme consists of a Borel probability measure
\[
\mu
\quad\text{on}\quad
\{-1,1\}^{\mathbb R^k}\times\{-1,1\}^{\mathbb R^k}.
\]
A sample \((f,g)\) from \(\mu\) provides Boolean functions \(f,g:\mathbb R^k\to\{-1,1\}\). Given vectors \(x_i,y_j\), the scheme preprocesses them into \(x_i',y_j'\), samples a Gaussian matrix \(G\), and outputs
\[
\varepsilon_i=f(Gx_i'),\qquad
\delta_j=g(Gy_j').
\]

The scheme has quality \(K\) when the transformed vectors can be chosen so that
\[
\mathbb E_G\left[
\int f(Gx_i')g(Gy_j')\,d\mu(f,g)
\right]
=
\frac{1}{K}\langle x_i,y_j\rangle
\]
for every \(i,j\). Multiplying by \(a_{ij}\), summing, and using that the outputs are signs yields the Grothendieck inequality with constant \(K\).

Gaussian rotational invariance implies that every scheme induces a scalar correlation function
\[
\Psi(t)=
\mathbb E_G\left[
\int f(Gx)g(Gy)\,d\mu(f,g)
\right],
\qquad \langle x,y\rangle=t.
\]
The preprocessing must realize transformed inner products satisfying
\[
\Psi(\langle x_i',y_j'\rangle)
=
\frac{1}{K}\langle x_i,y_j\rangle.
\]
When \(\Psi\) is invertible, the formal target is
\[
\langle x_i',y_j'\rangle
=
\Psi^{-1}\!\left(\frac{\langle x_i,y_j\rangle}{K}\right).
\]
Tensor-power maps provide a way to realize the required nonlinear inner products simultaneously.

The classical one-dimensional scheme uses
\[
f=g=\operatorname{sign}.
\]
For correlated standard Gaussians,
\[
\mathbb E[\operatorname{sign}(X)\operatorname{sign}(Y)]
=
\frac{2}{\pi}\arcsin(\langle x,y\rangle).
\]
The inverse correlation is \(\sin\). Tensorizing the odd Taylor expansion of \(\sin\) yields transformed unit vectors whose inner products equal
\[
\sin(c\langle x,y\rangle),
\qquad
c=\operatorname{arcsinh}(1)=\log(1+\sqrt2).
\]
Consequently,
\[
K_G^{\mathbb R}
\le
\frac{\pi}{2\log(1+\sqrt2)}
\approx 1.78221.
\]
This is the classical Krivine bound, not the exact value of \(K_G\). The bound is strictly larger than the real Grothendieck constant, so Krivine’s original conjecture that it was exact is false [1711.10595].

Naor and Regev proved that the apparent defect of the classical one-dimensional rule does not imply a defect of the general method. For every \(k\ge1\), there exists a \(k\)-dimensional oblivious Krivine scheme of quality at most
\[
\left(1+\frac{C}{k}\right)K_G,
\]
where \(C\) is universal. Their construction begins with a measure realizing the normalized inner product at scale \(1/K_G\), analyzes a Gaussian correlation function \(f_k\), and constructs nonlinear tensor maps from the inverse series of \(f_k\). The resulting quality satisfies
\[
K_k\le
\left(1+\frac{C}{k}\right)K_G,
\]
so high-dimensional oblivious Krivine schemes approximate \(K_G\) arbitrarily well [1205.6415].

The theorem does not establish Krivine’s original numerical conjecture. It establishes asymptotic optimality of the class of oblivious schemes when the dimension is allowed to grow.

## 4. Generalized Krivine rounding and stochastic variants

Krivine rounding extends beyond the sign-valued endpoint corresponding to \((p,q)=(\infty,1)\). For the \((\ell_p,\ell_q)\)-Grothendieck problem in the convex range
\[
p\ge2\ge q,
\]
the vector relaxation has the form
\[
\max\sum_{i,j}A_{ij}\langle u^i,v^j\rangle
\]
subject to
\[
\sum_i\|u^i\|_2^{q^*}\le1,
\qquad
\sum_j\|v^j\|_2^{p}\le1.
\]
The relaxation is convex precisely because \(p/2\ge1\) and \(q^*/2\ge1\).

The scalar sign map is replaced by Hölder-dual Gaussian powers:
\[
\operatorname{sgn}(w_i)|w_i|^{q-1}
\quad\text{and}\quad
\operatorname{sgn}(w_i)|w_i|^{p^*-1}.
\]
Writing
\[
a=p^*-1,\qquad b=q-1,
\]
the normalized Gaussian correlation is
\[
\widetilde f_{a,b}(\rho)
=
\rho\,
{}_2F_1\left(
\frac{1-a}{2},
\frac{1-b}{2};
\frac32;
\rho^2
\right).
\]
The classical arcsine correlation is the endpoint case \(a=b=0\).

As in the classical construction, the inverse correlation function is expanded into a power series and simulated by tensor powers. If
\[
\widetilde f_{a,b}^{-1}(\rho)=\sum_{k\ge1}c_k\rho^k,
\]
define the coefficientwise absolute majorant
\[
h_{a,b}(\rho)=\sum_{k\ge1}|c_k|\rho^k.
\]
The admissible scaling is
\[
c_{a,b}=h_{a,b}^{-1}(1).
\]
The resulting approximation guarantee is
\[
\frac{1+\epsilon_0}
{\sinh^{-1}(1)\gamma_{p^*}\gamma_q},
\qquad
\epsilon_0\le0.00863,
\]
with the equivalent \((p,r)\)-formulation obtained by setting \(q=r^*\). The analysis connects approximation ratios with Gaussian moments, hypergeometric functions, inverse Taylor coefficients, and factorization through Hilbert space [1804.03644].

A related stochastic formulation is given by Krivine diffusions. Let \(B(t)\) be Brownian motion and let
\[
dW_u^\varphi(t)
=
\varphi(W_u^\varphi(t))
\langle u,dB(t)\rangle,
\]
where \(\varphi\) is positive on \((-1,1)\) and vanishes at the endpoints. The terminal value is a sign
\[
\sigma_u^\varphi=\lim_{t\to\infty}W_u^\varphi(t).
\]
For the special coefficient
\[
\xi(s)
=
2\Phi'\!\left(
\Phi^{-1}\!\left(\frac{1-s}{2}\right)
\right),
\]
the terminal signs satisfy
\[
\mathbb E[\sigma_u^\xi\sigma_v^\xi]
=
\frac{2}{\pi}\arcsin(\langle u,v\rangle).
\]
Thus the diffusion reproduces exactly the Gaussian hyperplane-rounding kernel. Applied to MAXCUT, it attains the Goemans–Williamson ratio
\[
\rho_{\mathrm{GW}}
=
\min_{0\le\theta\le\pi}
\frac{2\theta/\pi}{1-\cos\theta}
\approx0.878567.
\]
The diffusion is therefore not a better MAXCUT approximation algorithm than Goemans–Williamson; its significance is the realization of the same kernel as the terminal law of a slowed-down diffusion [1906.10615].

Recent work has also investigated finite Gaussianized Hermite-chaos schemes. These use reservoirs based on Hermite polynomials of degrees such as \(3\) and \(5\), with correlations \(t^3\) and \(t^5\) rather than \(t\). Anti-alignment of the third-chaos component and alignment of the fifth-chaos component can suppress low-order nonlinear terms in the inverse correlation. The resulting improvement over the classical hyperplane radius is formulated through an inverse-majorant condition and transferred from an asymptotic weighted model to finite-dimensional schemes using locally uniform convergence, Rouche’s theorem, Cauchy estimates, and Wiener-algebra bounds. The advertised numerical conclusion in the supplied draft remains conditional on an unfinished interval certificate and contains an apparent inconsistency in translating the target radius into the final bound [2608.11158].

## 5. Krivine–Stengle schemes in polynomial optimization

Let
\[
\mathcal K
=
\{\mathbf x\in\mathbb R^n:
0\le g_i(\mathbf x)\le1,\ i=1,\ldots,p\}
\]
be a compact basic semialgebraic set. For multi-indices \(\boldsymbol\alpha,\boldsymbol\beta\in\mathbb N^p\), define
\[
h_{\boldsymbol\alpha,\boldsymbol\beta}(\mathbf x)
=
\prod_{i=1}^p
g_i(\mathbf x)^{\alpha_i}
(1-g_i(\mathbf x))^{\beta_i}.
\]
Every such generator is nonnegative on \(\mathcal K\).

The Krivine–Stengle positivity theorem states that if a polynomial \(q\) is strictly positive on \(\mathcal K\), then, for some \(k\),
\[
q(\mathbf x)
=
\sum_{|\boldsymbol\alpha+\boldsymbol\beta|\le k}
\lambda_{\boldsymbol\alpha,\boldsymbol\beta}
h_{\boldsymbol\alpha,\boldsymbol\beta}(\mathbf x),
\qquad
\lambda_{\boldsymbol\alpha,\boldsymbol\beta}\ge0.
\]
Restricting the total exponent to \(k\) defines a finite-order Krivine scheme. Increasing \(k\) enlarges the certificate cone.

For polynomial minimization,
\[
f^\star=\min_{\mathbf x\in\mathcal K}f(\mathbf x),
\]
the order-\(k\) relaxation is
\[
f_k^\star
=
\max\{t:f-t\text{ has a degree-}k\text{ Krivine--Stengle representation}\}.
\]
It satisfies
\[
f_k^\star\le f^\star,
\qquad
f_k^\star\longrightarrow f^\star.
\]
The relaxation is a linear program because the coefficients \(\lambda_{\boldsymbol\alpha,\boldsymbol\beta}\) enter linearly and polynomial identities are imposed by coefficient matching.

Dense representations become large rapidly. For a box with \(n\) coordinate constraints, the number of LP variables is
\[
\binom{2n+k}{k}+1,
\]
and the number of coefficient constraints is
\[
\binom{n+k}{k}
\]
when the constraint polynomials are linear.

Sparse Krivine–Stengle schemes use variable blocks \(I_j\) and constraint blocks \(J_j\). They require:

1. a decomposition \(f=\sum_j f_j\) with \(f_j\) depending only on variables in \(I_j\);
2. local constraints \(g_i\) depending only on variables in \(I_j\) for \(i\in J_j\);
3. coverage of all variables and constraints;
4. the running intersection property.

Under these assumptions, a positive polynomial can be decomposed as
\[
f=\sum_{j=1}^m\phi_j,
\qquad
\phi_j>0\text{ on }\mathcal K_j,
\]
where each \(\phi_j\) has a local Krivine–Stengle representation. The running intersection property preserves asymptotic convergence rather than merely reducing the number of variables.

In certified roundoff analysis, the floating-point model is
\[
\operatorname{rnd}(z)=z(1+e),
\qquad |e|\le\varepsilon.
\]
The roundoff error is decomposed as
\[
r(\mathbf x,\mathbf e)=l(\mathbf x,\mathbf e)+h(\mathbf x,\mathbf e),
\]
where \(l\) is the first-order error and \(h\) is bounded separately by interval arithmetic. After scaling by \(\varepsilon\), the first-order part is optimized over \([-1,1]^m\).

The sparse block for error variable \(e_j\) contains all input variables and only \(e_j\):
\[
I_j=\{1,\ldots,n,n+j\}.
\]
The corresponding local generators contain factors
\[
\left(\frac12+\frac{e_j}{2}\right)^{\gamma_j}
\left(\frac12-\frac{e_j}{2}\right)^{\delta_j}.
\]
This avoids products involving several distinct error variables.

The resulting lower and upper LP relaxations produce
\[
\underline l'_k\uparrow \underline l',
\qquad
\overline l'_k\downarrow \overline l',
\]
and therefore certified enclosures
\[
I_k^l=
[\varepsilon\underline l'_k,\varepsilon\overline l'_k].
\]
Adding an interval enclosure \(I^h\) of the higher-order remainder gives
\[
I_k=I_k^l+I^h.
\]
The associated implementations are FPBern and FPKriSten. The Krivine-based implementation is particularly suited to polynomial programs over semialgebraic domains and to problems with many error variables, whereas Bernstein methods are particularly effective for low-dimensional box-constrained problems and also handle rational programs [1610.07038] [1802.04385].

## 6. Krivine machines, automata, and realizability

A Krivine Abstract Machine evaluates simply typed or untyped lambda terms using closures, environments, and an argument stack. A closure is a pair \((\psi,e)\), consisting of an expression and the environment in which it is evaluated. Application pushes an argument closure onto the stack; variable lookup retrieves closures from environments; continuation mechanisms can capture and restore stacks.

Alternating Parity Krivine Automata combine this machinery with alternating parity tree automata. They operate on infinite full binary trees and provide operational semantics for Higher-Order Modal Fixpoint Logic. Their configurations contain a tree position, a current expression, environments, a closure stack, and a priority stack.

When a fixpoint state is entered, its arguments are stored in a fresh environment and its priority is pushed. When ground-type evaluation returns from an environment, the corresponding priority is removed. Acceptance is therefore governed not by ordinary parity on all visited states, but by a stair parity condition: only priorities that remain permanently on the dynamically evolving stack matter. An infinite play is accepting when the greatest priority that is never eventually popped is even.

APKA and HFL translate into one another without increasing type order. A state \(X\) becomes a fixpoint expression
\[
\sigma X.\lambda f_1^X\cdots\lambda f_{n_X}^X.\delta(X),
\]
where \(\sigma=\mu\) for odd priorities and \(\sigma=\nu\) for even priorities. Conversely, HFL formulas can be normalized into APKA states with typed arguments and priority assignments. The translations establish equal expressive power at each fixed order.

The number and parity of priorities induce a strict hierarchy. For the semantic classes \(\Sigma_n^{\mathrm{sem}}\) and \(\Pi_n^{\mathrm{sem}}\), the inclusions
\[
\Sigma_n^{\mathrm{sem}}
\subsetneq
\Sigma_{n+1}^{\mathrm{sem}},
\qquad
\Pi_n^{\mathrm{sem}}
\subsetneq
\Pi_{n+1}^{\mathrm{sem}}
\]
hold over the full class of infinite binary trees. The proof encodes acceptance games into trees and uses Arnold’s diagonal method together with Banach’s Fixpoint Theorem. The encoding map is a contraction under the metric
\[
d(\mathcal T,\mathcal T')=2^{-i},
\]
where \(i\) is the first level at which the trees differ. Its unique fixed point yields the self-referential tree required for the contradiction argument [1609.04092].

Krivine realizability uses related machine concepts but has a different semantic objective. A realizability algebra consists of closed lambda terms, stacks, a reduction relation, and a pole \(\Perp\) of successful processes. Orthogonality is defined by
\[
t\perp\pi
\quad\Longleftrightarrow\quad
t\star\pi\in\Perp.
\]
A term realizes a formula when it is orthogonal to every stack in the formula’s falsity value.

The control operator \(\mathsf{cc}\) captures the current continuation and realizes Peirce’s law,
\[
((\varphi\to\psi)\to\varphi)\to\varphi.
\]
This permits classical logic, unlike ordinary intuitionistic realizability. From a realizability algebra and a ground model of \(\mathsf{ZF}\), one constructs a universe of names and interprets formulas through truth and falsity values. Under coherence, the resulting theory is consistent; after extensional collapse, it yields a model of ordinary ZF. Special instructions such as `quote` support non-extensional choice principles and dependent choice. Boolean-valued models and forcing arise as special cases or limiting semantic forms of the general framework [2307.13563].

These computational meanings should not be conflated with Krivine rounding. A Krivine rounding scheme is a probabilistic or measurable transformation used in functional inequalities and approximation algorithms. A Krivine–Stengle scheme is a positivity-certificate hierarchy. A Krivine Abstract Machine is an operational evaluator. A Krivine realizability algebra is a machine-plus-pole semantics. Their common feature is the use of structured transformations—probabilistic, algebraic, or operational—to recover a target semantic property, but their mathematical objects and purposes are different.

Source: https://www.emergentmind.com/topics/krivine-schemes