Sharp Random Thresholds for Exact Ternary Homogeneous Boolean Lifts
Abstract
We study exact interpolation of Boolean functions by expanded homogeneous integer forms in the standard \({0,1}\)-encoding. A homogeneous form
is evaluated on the Boolean cube by restricting to the affine chart \(X_0=1\). We require all homogeneous coefficients of \(F\) to lie in \({-1,0,1}\), and we ask for exact equality
The evaluation map has an exact block structure: the number of homogeneous degree-\(d\) monomials collapsing to a Boolean monomial \(x_S\) is
Consequently, an exact ternary homogeneous lift of degree \(d\) exists if and only if every Boolean Möbius coefficient satisfies
For a uniformly random Boolean function
let \(D_{\mathrm{ex},1}\(G_n\)\) denote the least degree \(d\) for which an exact ternary homogeneous lift exists. We prove
where \(c>1\) is the unique solution of
and
Numerically,
The leading term and the \(\log n\) correction are forced by the top Boolean Möbius coefficient. The additional \(\log\log n\) correction is forced by the maximum of the \(n\) Möbius coefficients on the level (n-1). Thus the random exact-lift threshold is localized up to bounded fluctuations in probability.
The result concerns exact interpolation in a fixed expanded homogeneous integer presentation model. It is not a statement about ordinary polynomial threshold degree or about threshold sign-representation.
1. Introduction
Let
be uniformly random. We embed the Boolean cube into projective space by
A homogeneous form
is evaluated by
This note studies exact interpolation by such forms under the coefficient alphabet constraint
The model is deliberately presentation-dependent. It is not invariant under arbitrary Boolean recodings. Its advantage is that it admits a precise block decomposition. Each Boolean monomial \(x_S\) receives contributions from exactly
homogeneous monomials of degree \(d\). Hence the exact ternary lift problem becomes a family of independent capacity constraints on the Boolean Möbius coefficients.
The main result is a sharp random threshold theorem:
The leading constant \(c\) is determined by the top coefficient scale. Indeed, \(\mu_{[n]}\(G_n\)\) has standard deviation \(2^{n/2}\), while
where
Balancing the exponential scales gives
The correction \(\frac{\log n}{2L}\) compensates for the Stirling factor \(n^{-1/2}\). The further correction \(\frac{\log\log n}{2L}\) is caused by the level (n-1): there are \(n\) Möbius coefficients on that level, and their maximum after normalization is of order \(\sqrt{\log n}\).
2. Boolean Möbius coefficients
For \(S\subseteq[n]\), write
For an integer-valued function
define its Boolean Möbius coefficient by
where \(1_T\in{0,1}^n\) is the indicator vector of \(T\).
The unique multilinear representative of \(v\) is
We use the convention
3. Homogeneous block decomposition
Lemma 3.1. Block size
Fix \(S\subseteq[n]\), \(|S|=s\). The number of homogeneous monomials of degree \(d\) in \(X_0,\dots,X_n\) which reduce to \(x_S\) after substituting \(X_0=1\) and using \(x_i^k=x_i\) on the Boolean cube is
Proof
A monomial
of total degree \(d\) reduces to \(x_S\) if and only if
and
Writing
we count nonnegative solutions of
There are (s+1) variables, so the number of solutions is
\(\square\)
4. Exact ternary lifts
Let
Define \(D_{\mathrm{ex},1}(g)\) to be the least integer \(d\) such that there exists
with all homogeneous coefficients in \({-1,0,1}\), satisfying
Lemma 4.1. Exact ternary lift criterion
There exists an exact homogeneous degree-\(d\) lift of \(g\) with coefficients in \({-1,0,1}\) if and only if
Consequently,
Proof
By Lemma 3.1, the coefficient of \(x_S\) in the multilinear representative of (F(1,x)) is the sum of exactly \(\binom d{|S|}\) homogeneous coefficients. If each homogeneous coefficient lies in \({-1,0,1}\), then this sum can be any integer in
Since the coefficient of \(x_S\) in the multilinear representative of \(g\) is \(\mu_S(g)\), the displayed inequalities are necessary.
Conversely, if
for every \(S\), then in the block corresponding to \(S\) choose \(|\mu_S(g)|\) coefficients equal to \(\operatorname{sgn}\mu_S(g)\), and set the remaining coefficients in the block equal to (0). The blocks are disjoint, so these choices can be made independently for all \(S\). The resulting homogeneous form has coefficients in \({-1,0,1}\) and evaluates exactly to \(g\). \(\square\)
5. Random Möbius coefficients
Let
be uniformly random, with independent values.
For fixed \(S\), \(|S|=s\), the coefficient
is a sum of \(2^s\) independent Rademacher random variables. Hence
Define the normalized capacity
Hoeffding's inequality gives
6. Constants and binomial asymptotics
Let
Then
Thus there is a unique \(c>1\) satisfying
Set
Numerically,
and
Define
Lemma 6.1. Top binomial asymptotic
Uniformly for \(u=O(1)\),
More generally, uniformly for \(u=O\(\log n\)\),
Proof
Stirling's formula gives, uniformly for \(d=cn+O\(\log n\)\),
Writing \(d=cn+u\), with \(u=O\(\log n\)\), and expanding at \(c\), we get
Since \(f'(c)=L\), this gives
Since \(f(c)=\frac12\log2\), substituting
yields the first formula. \(\square\)
Lemma 6.2. Near-top ratio
Let \(d=cn+O\(\log n\)\). For \(0\le k\le n\),
In particular, for sufficiently small fixed \(\delta>0\), there exists \(a>1\) such that for all \(0\le k\le\delta n\), and all sufficiently large \(n\),
Proof
We compute
Since
the identity follows.
If \(d=cn+O\(\log n\)\), then for \(0\le k\le\delta n\),
For sufficiently small \(\delta>0\),
This gives the claim. \(\square\)
7. Main theorem
Theorem 7.1
Let \(G_n:{0,1}^n\to{-1,1}\) be uniformly random. Then
Equivalently, the sequence
is tight.
8. Upper bound
We prove that there exists \(M<\infty\) such that
Let
By Lemma 4.1 and Hoeffding's inequality,
We split the levels into three ranges.
Low levels
Choose \(\eta>0\) sufficiently small so that
For \(1\le s\le\eta n\), and all sufficiently large \(n\),
But deterministically
Thus no failure can occur in this range. The case \(s=0\) is also deterministic:
Middle levels
Fix \(\delta>0\). For
write \(s=\rho n\). Stirling's formula gives, uniformly for \(\rho\in[\eta,1-\delta]\),
Since \(\rho<1\), we have \(c/\rho>c\). Because \(f\) is strictly increasing and
the bracket is bounded below by a positive constant depending only on \(\eta,\delta\). Hence
uniformly in the middle range, for some \(\kappa>0\). Therefore the total middle-level contribution is at most
Near-top levels
Let
where \(\delta>0\) is chosen sufficiently small for Lemma 6.2.
By Lemmas 6.1 and 6.2,
where \(a>1\) is fixed and \(C_M\to\infty\) as \(M\to\infty\).
Choose \(M\) sufficiently large so that
Then
Summing over \(0\le k\le\delta n\), the near-top contribution is (o(1)).
Combining all ranges gives
Therefore
with probability tending to (1).
9. Lower bound
We prove that there exists \(M<\infty\) such that
Let
We use only the level (n-1). For \(i=1,\dots,n\), set
and
We show that
with probability tending to (1), for some absolute constant \(c_0>0\).
Define
The variables \(\eta_T\) are independent Rademacher variables.
For \(i=1,\dots,n\), define
Set
and
Since
we have
Let
Then
where
The vectors \(Y_T\) are independent, mean zero, and
because the Walsh characters \(\chi_i\) are orthogonal on the Boolean cube.
Moreover,
Therefore
We use the following standard multivariate Berry--Esseen estimate for convex sets: if independent mean-zero random vectors in \(\mathbb{R}^m\) have total covariance \(I_m\), then
where \(\mathcal C_m\) is the class of convex Borel subsets of \(\mathbb{R}^m\), \(Z\sim N\(0,I_m\)\), and \(C_0\) is an absolute constant.
In our case, \(m=n\), so the error is at most
Therefore, for every rectangle
we have
where \(Z=\(Z_1,\dots,Z_n\)\) is a standard Gaussian vector in \(\mathbb{R}^n\).
Take
Then
Since
this probability tends to (0). Hence
On the other hand, \(A\) is a normalized Rademacher sum, so Hoeffding gives
Consequently, with probability tending to (1),
Since
we obtain
with probability tending to (1).
Now estimate the normalized capacity at level (n-1). By Lemmas 6.1 and 6.2,
where
Choose \(M\) sufficiently large so that
Then, with probability tending to (1), there exists \(i\) such that
By Lemma 4.1, no exact ternary homogeneous lift of degree \(d_n^-\) exists. Hence
with probability tending to (1).
Thus
with high probability.
Combining the upper and lower bounds proves Theorem 7.1. \(\square\)
10. Interpretation
The first-order constant \(c\) comes from balancing
against the natural scale \(2^{n/2}\) of the top Möbius coefficient.
The correction
compensates for the Stirling factor \(n^{-1/2}\).
The additional correction
is forced by the \(n\) coefficients
After normalization, these coefficients have maximum of order \(\sqrt{\log n}\). This factor translates into the additive degree shift
Thus the top coefficient determines the (cn) and \(\log n\) terms, while the level (n-1) forces the \(\log\log n\) term.
11. Relation to threshold sign-representation
The exact interpolation problem studied above is distinct from threshold sign-representation.
In threshold sign-representation, one seeks an integer-valued function \(v\) such that
and then asks whether the Möbius coefficients of \(v\) fit inside the same homogeneous coefficient capacities. This additional freedom can substantially reduce the necessary degree.
We record the real dual formulation of the bounded-coefficient threshold problem, since it clarifies the distinction.
Let \(B_S\ge0\) be coefficient bounds, and define
Let \(\Delta\) be the simplex of probability measures on \(2^{[n]}\).
Then minimax duality gives
For the degree-\(n\), height-(1) homogeneous model,
The choice \(\lambda=\delta_\varnothing\) gives value (1), so
for every \(g\).
Thus the random threshold sign-representation problem asks whether, for random \(G_n\),
with high probability, and then whether a suitable integer rounding with slack is possible. This is not solved here.
A useful deterministic estimate for the RMS scale of the dual expression is the following.
Proposition 11.1. Square-root zeta lower bound
Let \(n\ge4\), and let \(\lambda\) be a probability measure on \(2^{[n]}\). Define
Then
In particular,
Proof
For nonempty \(S\), set
with \(m_S=0\) if there is no such \(T\). Since
we have
For \(t>0\), define
Then
where
We use the elementary fact that if
is a nonempty downset and \(n\ge4\), then
Indeed, if \([n]\notin\mathcal D\), then every \(S\in\mathcal D\) has \(1\le |S|\le n-1\), and \(\binom n{|S|}\ge n\). If \([n]\in\mathcal D\), then \(\mathcal D=2^{[n]}\setminus{\varnothing}\), and
for \(n\ge4\); the surplus at level (2) compensates the deficit at the top level.
Applying this to \(\downarrow\mathcal A_t\), we obtain
But
Therefore the nonempty part contributes at least
The \(S=\varnothing\) term is
This proves the first inequality. The second follows from
\(\square\)
This proposition shows that any dual measure with noticeable mass away from the origin has a large RMS zeta profile. It is included only as context for the harder threshold sign-representation problem.
12. Scope and limitations
This paper concerns exact interpolation by expanded homogeneous integer forms in one fixed presentation model. It does not give coordinate-invariant lower bounds for polynomial threshold functions.
The theorem does not solve the random bounded-height threshold sign-representation problem. In particular, it does not determine
or the random threshold degree
The main contribution is the sharp random degree threshold for exact ternary homogeneous lifts.
References
P. Baldi and R. Vershynin, Polynomial threshold functions, hyperplane arrangements, and random tensors, SIAM Journal on Mathematics of Data Science 1 (2019), no. 4, 699--729.
V. Bentkus, On the dependence of the Berry--Esseen bound on dimension, Journal of Statistical Planning and Inference 113 (2003), 385--402.
S. Boucheron, G. Lugosi, and P. Massart, Concentration Inequalities: A Nonasymptotic Theory of Independence, Oxford University Press, 2013.
R. O'Donnell, Analysis of Boolean Functions, Cambridge University Press, 2014.
R. O'Donnell and R. A. Servedio, New degree bounds for polynomial threshold functions, Combinatorica 30 (2010), 327--358.