Robust Daisies in RLDC Analysis

• The paper introduces the robust daisy framework, a probabilistic generalization of sunflower structures that enables nearly optimal lower bounds for linear q-query RLDCs.
• It leverages a flexible kernel selection, allowing for arbitrary overlapping petals and high probability coverage via binomial sampling.
• Key techniques such as spreadness and the small-set spread lemma connect combinatorial structures with pseudorandom properties, advancing RLDC analysis.

A robust daisy is a probabilistic generalization of the classical sunflower structure for set systems, designed to quantitatively capture the combinatorial and pseudorandom properties required in the analysis of relaxed locally decodable codes (RLDCs). Unlike ordinary daisies and sunflowers—which impose static kernel and petal structures—a robust daisy introduces the flexibility of a small kernel, arbitrary (possibly overlapping) petals, and distributional guarantees that allow for satisfaction via biased sampling, even after conditioning on large subfamilies. This concept, together with spreadness and associated extraction lemmas, underpins the nearly optimal lower bounds for the length of linear $q$-query RLDCs, closing the gap to long-standing upper bounds. The robust daisy framework also forges a connection between sunflower-type arguments and pseudorandom set system structure (Goldberg et al., 26 Nov 2025).

1. Formal Definition of Robust Daisies

Let %%%%1%%%% be an $n$-element universe and let $\mu$ be a probability distribution over the family of sets $\mathcal{P}_{\le q}(U)$, the subsets of $U$ of cardinality at most $q$ (the query complexity of the code). For a given set $K \subseteq U$ (the kernel), $\mu$ is a $(p,\varepsilon)$–robust daisy with kernel $K$ if, for every subfamily $D \subseteq \mathrm{supp}(\mu)$,

$$\Pr_{W \sim \mathrm{Bin}(U,p)}\left[\exists\,S \in D\quad S \subseteq K \cup W\right] \ge 1 - (\varepsilon)^{\mu(D)},$$

where $W$ is a binomial subset (each element of $U$ kept independently with probability $p$), and $\mu(D) = \sum_{S \in D} \mu(S)$. Thus, after “handling” the kernel coordinates, the probability that a random sample covers a set from $D$ is exponentially high in $\mu(D)$.

In contrast to sunflowers, the kernel in a robust daisy can be any small set, the distribution over query sets is arbitrary, and the petals may overlap. The crucial requirement is high probability “hitting” under binomial sampling, for every subfamily $D$—not only the entire system.

2. Spreadness: Pseudorandom Combinatorial Structure

Spreadness abstracts the requirement that no small subset of coordinates appears disproportionately in the support of $\mu$. Formally, $\mu$ is $(m,k)$–spread if for every nonempty $T \subseteq U$,

$$\mu(\langle T\rangle) = \sum_{S \supseteq T} \mu(S) \le \frac{m}{k^{|T|}},$$

where $\langle T\rangle$ is the star of $T$. Spreadness ensures that individual elements, pairs, or larger tuples appear only a small fraction of the time (bounded by $m/k^{|T|}$), generalizing the bounded-intersection property of daisies. Notably, spreadness is preserved under conditioning: if $\mu$ is $(m,k)$–spread and $D$ is a subfamily with $\mu(D) > 0$, then the conditional $\mu_D$ is $(m/\mu(D),k)$–spread.

This property enables robust analysis of subfamilies and connects the analytic perspective of distributions with traditional combinatorial sunflower bounds.

3. The Small–Set Spread Lemma

A core technical lemma is that $(m,k)$–spread distributions over small sets (of size at most $q$) are themselves “satisfying” under binomial sampling, that is, with high probability a random subset hits some set in the support. The precise statement is:

Small–Set Spread Lemma:

Given $q\ge1$, $m\in(0,1]$, and any $(m,k)$–spread distribution $\mu$ over subsets of size at most $q$, let $\alpha > 2q$, set $p = \alpha/k$ and $\varepsilon = \exp(-\Omega(\alpha/(q m)))$. Then

$$\Pr_{W \sim \mathrm{Bin}(U,p)}\left[\exists\, S \in \mathrm{supp}(\mu): S \subseteq W\right] \ge 1 - \varepsilon.$$

The proof leverages the expectation $\mathbb{E}[X]$ of the count $X$ of subsets hit, and uses Janson’s inequality with overlap control provided by the spread property. This lemma provides the quantitative ingredient for constructing robust-daisy properties with explicit parameters.

4. Robust Daisy Extraction from Arbitrary Distributions

For general $\mu$ (not initially spread), a structural extraction lemma demonstrates that by puncturing the universe on a small set $K$ and restricting to a large-mass subfamily $D$, the conditioned and punctured distribution $\mu_D^{\circ K}$ is $(m,k)$–spread with parameters $k \approx n^{1/c}$ and $m\,|K| = O(1)$ for $c > q$. Formally,

Spreadness Extraction Lemma:

For $\mu$ over $\mathcal{P}_{\le q}(U)$, $|U|=n$, and integer $c>q$, there exists $D \subseteq \mathrm{supp}(\mu)$ with $\mu(D)\ge 1 - q/c$, and kernel $K$ with $|K| \le n^{1-1/c}$, such that $\mu_D^{\circ K}$ is $(m, k)$–spread with $k = n^{1/c}$ and $m|K| = O(1)$.

This is achieved through an analysis of weighted degrees, degree bucketing of $U$, and token shifting, ensuring that post-puncturing and conditioning, the distribution concentrates on “good” subfamilies with the required spread property.

The synthesis of the spreadness extraction lemma and the small–set spread lemma yields the Robust Daisy Lemma, guaranteeing for any $\mu$ the existence of subfamily $D$ and kernel $K$ giving a $(p,\varepsilon)$–robust daisy structure with $p = \alpha/n^{1/c}$ and $\varepsilon = \exp(-\Omega(\alpha/q^2) |K|)$, for $\alpha > 2q$.

5. Application: Nearly Tight Lower Bounds for RLDCs

The robust daisy framework is applied to obtain nearly optimal lower bounds for the block length of linear $q$-query RLDCs. If a code $C: \{0,1\}^k \to \Sigma^n$ admits—per message index $i$—a query distribution $\mu_i$ that is a $(p,\varepsilon_i)$–robust daisy with kernel $K_i$ of size $|K_i| \le \delta n$ and $\varepsilon_i^{\sigma} \le 1/(3k|\Sigma|^{|K_i|})$, then a global sampler $G$ can simultaneously recover all message bits with only $O(p n)$ queries:

• $G$ samples $W \sim \mathrm{Bin}([n],p)$, querying these positions.
• For each $i$ and each candidate assignment $\kappa$ to $K_i$, $G$ checks all relevant “petals,” applies local decoding, and seeks unanimity.

The concentration principle and robust daisy property ensure successful decoding with high probability. This leads to an upper bound for $k$ in terms of $n$: $k \le O(p n \log |\Sigma|)$. Choosing parameters appropriately and applying the robust daisy extraction gives $n \ge k^{1+\Omega(1/q)}$, matching the best-known upper bounds $n = k^{1+O(1/q)}$ up to constants (Goldberg et al., 26 Nov 2025).

6. Relation to Sunflowers and Daisies

In classical sunflower lemmas, all sets share a static intersection kernel and their petals are disjoint. Robust daisies generalize this setup by relaxing kernel selection, allowing arbitrary overlaps, and employing a distributional perspective. Sunflower structures are recovered when the distribution is uniform and the kernel is the intersection of all sets. The robust daisy is thus a relaxation—suitable when the set system under study is inherently pseudorandom or structured irregularly.

The key distinction is that robust daisies are defined by satisfaction under binomial sampling for all subfamilies after kernel removal, versus the blanket disjointness and intersection constraints of sunflowers and classical daisies.

7. Implications and Significance

The robust daisy framework provides a unified approach to analyzing code locality, combining structural and probabilistic combinatorics. It resolves a major open question by proving the nearly tight lower bound $n \ge k^{1+\Omega(1/q)}$ for linear $q$–query RLDCs, as previously several techniques failed to match the best upper bound except in special cases. This suggests that robust daisies may be a fundamental object for understanding relaxed coding structures, with potential applications to other pseudorandom or approximate combinatorial scenarios. Robust daisies clarify the analogy between sunflowers in set systems and “hitting” properties needed for relaxed decoding and may inform future developments in both coding theory and combinatorial design (Goldberg et al., 26 Nov 2025).