Introduction
The question of graph coloring , which colors each vertex of a given graph G = ( V , E ) G=(V,E) G = ( V , E ) with the smallest number of colors so that adjacent vertices always have different colors, is a central topic in both discrete math and theoretical computer science. This is not only one of the most important NP-hard problems, but even deciding whether three colors suffice is among Karp’s 21 classic NP-complete problems [23 , 34 ] .
Indeed, the fastest known algorithm to solve graph coloring takes O ∗ ( 2 n ) O^*(2^n) O ∗ ( 2 n ) time [6 ] , where n n n is the number of vertices. For determining whether the graph is k k k -colorable, the fastest known algorithm takes O ∗ ( 1.3289 n ) O^*(1.3289^n) O ∗ ( 1.328 9 n ) time for k = 3 k=3 k = 3 [4 ] , O ∗ ( 1.7159 n ) O^*(1.7159^n) O ∗ ( 1.715 9 n ) time for k = 4 k=4 k = 4 [49 ] , and O ∗ ( ( 2 − ε k ) n ) O^*((2-\varepsilon_k)^n) O ∗ (( 2 − ε k ) n ) time for k = 5 , 6 k=5,6 k = 5 , 6 for some very small ε k > 0 \varepsilon_k>0 ε k > 0 [51 ] . Very recently, two independent preprints that claim O ∗ ( ( 2 − ε k ) n ) O^*((2-\varepsilon_k)^n) O ∗ (( 2 − ε k ) n ) time for all k ∈ N k\in\mathbb{N} k ∈ N appeared [Zam26; Pra26]. Polynomial-time algorithms are known only for limited graph classes, including perfect graphs [26 ] and bounded-treewidth graphs [CFK+15].
Approximation Algorithms
Approximation algorithms are well studied for NP-hard problems in general. However, graph coloring is NP-hard even to approximate within an n 1 − ε n^{1-\varepsilon} n 1 − ε factor for all ε > 0 \varepsilon>0 ε > 0 [20 , 31 , 53 ] . Note that the best known approximation factor is O ( n ( log log n ) 2 / ( log n ) 3 ) O(n(\log\log n)^2/(\log n)^3) O ( n ( log log n ) 2 / ( log n ) 3 ) [13 , 30 ] .
Given very strong negative results for the approximation, the following question became the main line of research: “Given a k k k -colorable graph G = ( V , E ) G=(V,E) G = ( V , E ) , find a coloring in polynomial time, using a sufficiently small number of colors.” In particular, the case k = 3 k=3 k = 3 is important for being the smallest nontrivial case, and improving this case typically yields a better algorithm for all subsequent k ≥ 4 k\geq4 k ≥ 4 [12 , 32 ] .
This topic also connects deeply to a central problem in complexity theory, the Unique Games Conjecture , and its variant, the d-to-1 Conjecture , both introduced by Khot [35 ] . If the d-to-1 Conjecture is true for some constant d d d , then it is NP-hard to color 3-colorable graphs with any constant number of colors in polynomial time [18 , 28 ] . Very recently, a preprint that claims to resolve the 4-to-1 Conjecture appeared [FMW26], which, if true, would resolve the impossibility for any constant number of colors. Meanwhile, unconditional NP-hardness was already known for coloring k k k -colorable graphs with 2 k − 1 2k-1 2 k − 1 colors [36 , 24 , 9 ] .
Polynomial-Time Coloring of 3-Colorable Graphs
Wigderson [50 ] initiated research on coloring 3-colorable graphs in 1982, presenting an O ( n ) O(\sqrt{n}) O ( n ) -coloring algorithm. Since then, researchers have made a long line of improvements over 40 years, approaching the problem from two sides: the large-degree case and the small-degree case. We present the sequence of improvements in Table 1.
Two directions of the attack. Generally speaking, if we want to create an algorithm to color 3-colorable graphs with O ~ ( n α ) \widetilde{O}(n^\alpha) O ( n α ) colors (for α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) ), we consider an algorithm to make “progress” toward O ~ ( n α ) \widetilde{O}(n^\alpha) O ( n α ) -coloring. Here, the progress includes obtaining an independent set of size Ω ~ ( n 1 − α ) \widetilde{\Omega}(n^{1-\alpha}) Ω ( n 1 − α ) , as repeating the routine of “coloring the independent set with a common new color and removing these vertices from the graph” until the graph becomes empty gives an O ~ ( n α ) \widetilde{O}(n^\alpha) O ( n α ) -coloring. Note that, while Blum [12 ] also considers other types of progress,5 which are used in solving the large-degree case including [42 ] , we do not take them into account in this paper.
Table 1. Table 1: The line of research for coloring 3-colorable graphs in polynomial time. Although not listed below, some very recent preprints claim the following results: O ~ ( n 0.1751 ) \widetilde{O}(n^{0.1751}) O ( n 0.1751 ) colors on September 27, 2026 [46 ] , O ~ ( n 0.1780 ) \widetilde{O}(n^{0.1780}) O ( n 0.1780 ) colors on September 30, 2026 [45 ] , O ~ ( n 0.1740 ) \widetilde{O}(n^{0.1740}) O ( n 0.1740 ) colors on October 1, 2026 [3 ] , and O ~ ( n 0.1653 ) \widetilde{O}(n^{0.1653}) O ( n 0.1653 ) colors on October 2, 2026 [48 ] .
The following framework is widely used: for an arbitrary threshold Δ \Delta Δ , it suffices to consider making progress toward O ~ ( n α ) \widetilde{O}(n^\alpha) O ( n α ) -coloring for graphs with maximum degree ≤ Δ \le\Delta ≤ Δ and for graphs with minimum degree ≥ Δ \ge\Delta ≥ Δ ([41 ] , Proposition 7.1). For this reason, an improvement on either the large-degree case or the small-degree case yields a better algorithm for 3-colorable graphs.
The starting point: Wigderson’s algorithm. For graphs with minimum degree Δ \Delta Δ , the induced subgraph G [ N ( x ) ] G[N(x)] G [ N ( x )] on the neighborhood of x ∈ V x \in V x ∈ V is bipartite, with at least Δ \Delta Δ vertices. Hence, an independent set of size ≥ Δ / 2 \ge\Delta/2 ≥ Δ/2 can be obtained in linear time (by bipartite coloring), making progress toward O ( n / Δ ) O(n/\Delta) O ( n /Δ ) -coloring. Graphs with maximum degree Δ \Delta Δ can be colored with Δ + 1 \Delta+ 1 Δ + 1 colors by greedy coloring. Wigderson combined them with threshold Δ = n \Delta= \sqrt{n} Δ = n and yielded an O ( n ) O(\sqrt{n}) O ( n ) -coloring algorithm.
Large-degree case and combinatorics. In 1989, Blum achieved the first improvement over Wigderson’s algorithm; his line of work [Blu89; Blu90; Blu94] yields a better algorithm for the large-degree cases, which makes progress toward O ~ ( ( n / Δ ) 3 / 5 ) \widetilde{O}((n/\Delta)^{3/5}) O (( n /Δ ) 3/5 ) -coloring. The attack on the large-degree case is mainly done in a combinatorial way, analyzing dense substructures of a graph. The best known algorithm for the large-degree case makes progress toward ( n / Δ ) 1 / 2 n o ( 1 ) (n/\Delta)^{1/2}n^{o(1)} ( n /Δ ) 1/2 n o ( 1 ) -coloring for every fixed ε > 0 \varepsilon> 0 ε > 0 , whenever the minimum degree Δ > n \Delta> \sqrt{n} Δ > n [42 ] . Therefore, we can use the following tool when combined with the small-degree case.
Theorem 1 ([42 ] , restated). Fix α ∈ ( 0 , 1 4 ] \alpha\in(0, \frac{1}{4}] α ∈ ( 0 , 4 1 ] . Suppose that there exists an algorithm that, given a 3-colorable graph with maximum degree Δ < n 1 − 2 α \Delta< n^{1-2\alpha} Δ < n 1 − 2 α , obtains an independent set of size n 1 − α − o ( 1 ) n^{1-\alpha-o(1)} n 1 − α − o ( 1 ) in polynomial time. Then, one can color a 3-colorable graph with n α + o ( 1 ) n^{\alpha+o(1)} n α + o ( 1 ) colors in polynomial time.
Small-degree case and SDP. On the other hand, improvements on the small-degree case generally involve semidefinite programming (SDP) . Goemans and Williamson’s 1994 landmark result, which achieves a 0.878-approximation of the maximum cut problem via SDP [29 ] , sparked interest in applying SDP to other optimization problems. In the same year, Karger, Motwani, and Sudan [37 , 38 ] presented an SDP-based algorithm that, given a 3-colorable graph with maximum degree Δ \Delta Δ , computes an O ~ ( Δ 1 / 3 ) \widetilde{O}(\Delta^{1/3}) O ( Δ 1/3 ) -coloring in polynomial time. Their SDP formulation is “vector k k k -coloring ”, defined to be vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V in R n \mathbb{R}^n R n such that ∥ v x ∥ = 1 \lVert v_x\rVert= 1 ∥ v x ∥ = 1 for all x ∈ V x \in V x ∈ V and ⟨ v x , v y ⟩ ≤ − 1 k − 1 \langle v_x, v_y\rangle\le-\frac{1}{k-1} ⟨ v x , v y ⟩ ≤ − k − 1 1 for all x y ∈ E xy \in E x y ∈ E . It is not hard to see that if the graph is k k k -colorable, then a vector k k k -coloring exists and can be computed in polynomial time via a polynomial-time SDP algorithm [25 ] . A rounding scheme on vector 3-coloring enabled them to obtain an independent set of size Ω ~ ( n / Δ 1 / 3 ) \widetilde{\Omega}(n/\Delta^{1/3}) Ω ( n / Δ 1/3 ) with high probability.
Combining this with [42 ] ’s framework (Theorem 1) gives n 1 / 5 + o ( 1 ) n^{1/5+o(1)} n 1/5 + o ( 1 ) colors, splitting at Δ = n 3 / 5 \Delta= n^{3/5} Δ = n 3/5 . As further studies on the SDP side [1 , 17 , 7 ] improve the bounds when Δ \Delta Δ is large, the best known algorithm [7 ] achieves an exponent slightly better than 1 5 \frac{1}{5} 5 1 .
Related works. Some studies consider coloring 3-colorable graphs with a specific structure. Alon and Kahale [2 ] showed, using spectral graph theory, a polynomial-time algorithm to find a 3-coloring for random 3-colorable graphs. Recently, Hsieh [33 ] presented a polynomial-time algorithm to find a large independent set for 3-colorable graphs with low threshold rank.
Our Results
In this paper, we make an improvement to the small-degree cases for the problem of coloring 3-colorable graphs in polynomial time. Specifically, we obtain the following results.
Theorem 2. Fix a real number k > 2 k > 2 k > 2 and β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] . Let G = ( V , E ) G = (V,E) G = ( V , E ) be a graph on n n n vertices with maximum degree at most n β n^\beta n β , given together with an ordinary vector k k k -coloring, that is, unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V with
⟨ v x , v y ⟩ ≤ − 1 k − 1 \langle v_x,v_y\rangle\le-\frac{1}{k-1} ⟨ v x , v y ⟩ ≤ − k − 1 1 for every edge x y ∈ E xy \in E x y ∈ E . For every fixed C > 0 C > 0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , returns an independent set I I I with
∣ I ∣ ≥ { n 1 − k − 2 k β − o ( 1 ) , 0 < β ≤ k − 2 k , n ( k − 2 ) β 2 + k β − o ( 1 ) , k − 2 k < β ≤ 1. |I| \ge
\begin{cases}
n^{1-\frac{k-2}{k}\beta-o(1)}, & 0<\beta\le\frac{k-2}{k},\\
n^{\frac{(k-2)\beta}{2+k\beta}-o(1)}, & \frac{k-2}{k}<\beta\le1.
\end{cases} ∣ I ∣ ≥ { n 1 − k k − 2 β − o ( 1 ) , n 2 + k β ( k − 2 ) β − o ( 1 ) , 0 < β ≤ k k − 2 , k k − 2 < β ≤ 1.
Theorem 3. Fix β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] . Let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β , given together with a strict vector 3-coloring, that is, unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V with ⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1 for every edge x y ∈ E xy\in E x y ∈ E . For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , returns an independent set I I I with
∣ I ∣ ≥ { n 1 − β / 3 − o ( 1 ) , 0 < β ≤ 1 / 3 , n 2 + 2 β 2 + 3 β − o ( 1 ) , 1 / 3 < β ≤ 1. |I| \ge
\begin{cases}
n^{1-\beta/3-o(1)}, & 0<\beta\le1/3,\\
n^{\frac{2+2\beta}{2+3\beta}-o(1)}, & 1/3<\beta\le1.
\end{cases} ∣ I ∣ ≥ { n 1 − β /3 − o ( 1 ) , n 2 + 3 β 2 + 2 β − o ( 1 ) , 0 < β ≤ 1/3 , 1/3 < β ≤ 1.
Note that Theorem 2 achieves better performance than the algorithm of Karger, Motwani, and Sudan [38 ] when β > k − 2 k \beta>\frac{k-2}{k} β > k k − 2 and the same performance when 0 < β ≤ k − 2 k 0<\beta\le\frac{k-2}{k} 0 < β ≤ k k − 2 . The main ideas behind Theorem 2 already appear in the case k = 3 k=3 k = 3 . We therefore first prove Theorem 3, which assumes a strict vector 3-coloring. In Appendix § 8 , we extend the result from strict vector 3-colorings to ordinary vector k k k -colorings for every real k > 2 k>2 k > 2 and thereby prove Theorem 2 . Note that Theorem 2 does not require the parameter k k k to be an integer.
Since a strict vector k k k -coloring can be computed in polynomial time by SDP [38 ] , by combining Theorem 3 with [42 ] 's framework (Theorem 1), we obtain the following result.
Theorem 4. For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , properly colors every 3-colorable n n n -vertex graph with n 1 / 6 + o ( 1 ) n^{1/6+o(1)} n 1/6 + o ( 1 ) colors.
Proof. Given a 3-colorable graph with maximum degree at most n 2 / 3 n^{2/3} n 2/3 , we compute a strict vector 3-coloring by SDP (Subsection 3.1), and apply Theorem 3 with β = 2 3 \beta=\frac{2}{3} β = 3 2 . For any fixed C ′ > 0 C'>0 C ′ > 0 , this returns an independent set of size n 5 / 6 − o ( 1 ) n^{5/6-o(1)} n 5/6 − o ( 1 ) in polynomial time with probability at least 1 − n − C ′ 1-n^{-C'} 1 − n − C ′ . Thus, we can apply Theorem 1 with α = 1 6 \alpha=\frac{1}{6} α = 6 1 and obtain a coloring with n 1 / 6 + o ( 1 ) n^{1/6+o(1)} n 1/6 + o ( 1 ) colors. Since we call the independent-set algorithm only polynomially many times, choosing a sufficiently large C ′ > C C'>C C ′ > C guarantees the success probability of 1 − n − C 1-n^{-C} 1 − n − C . □ \square □
Moreover, we further strengthen the SDP algorithm specifically for 3-colorable graphs (beyond vector 3-colorable graphs) to obtain a larger independent set. We prove the following theorem.
Theorem 5. Let β = 0.692601 \beta=0.692601 β = 0.692601 . Let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β . For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , returns an independent set I I I with ∣ I ∣ ≥ n ( 1 + β ) / 2 − o ( 1 ) |I|\ge n^{(1+\beta)/2-o(1)} ∣ I ∣ ≥ n ( 1 + β ) /2 − o ( 1 ) .
Combining Theorem 5 with [42 ] 's framework (Theorem 1), we obtain the following result.
Theorem 6. For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , properly colors every 3-colorable n n n -vertex graph with O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) colors.
Proof. By Theorem 5, we can apply Theorem 1 with α = 1 − β 2 = 1 − 0.692601 2 < 0.1537 \alpha=\frac{1-\beta}{2}=\frac{1-0.692601}{2}<0.1537 α = 2 1 − β = 2 1 − 0.692601 < 0.1537 , and obtain a coloring with O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) colors. Since we call the independent-set algorithm only polynomially many times, choosing a sufficiently large C ′ > C C'>C C ′ > C guarantees the success probability of 1 − n − C 1-n^{-C} 1 − n − C . □ \square □
New Bounds on Vector k k k -Colorable Graphs
We recall that our result (Theorem 2 ) can be applied not only to k k k -colorable graphs but also to vector k k k -colorable graphs , that is, graphs that have (ordinary) vector k k k -colorings.
Lower bounds. While k k k -colorable graphs are always vector k k k -colorable, the converse may not hold, and their chromatic numbers can be arbitrarily large. Feige, Langberg, and Schechtman [21 ] give the following lower bound. In particular, there exist vector 3 3 3 -colorable graphs with chromatic number Ω ( n 0.1578 ) \mathit{\Omega}(n^{0.1578}) Ω ( n 0.1578 ) .
Theorem 7 ([21 ] , restated). Let k > 2 k > 2 k > 2 be a real constant. For infinitely many n n n , there exists an n n n -vertex vector k k k -colorable graph with chromatic number n α ( k ) − o ( 1 ) n^{\alpha(k)-o(1)} n α ( k ) − o ( 1 ) , where
α ( k ) = max 0 < a < k − 2 2 ( k − 1 ) log ( 1 − a 2 ) log { ( 1 − 1 ( k − 1 ) 2 ) ( 1 − 2 ( k − 1 ) k − 2 a 2 ) } − log ( 1 − a 2 ) . \alpha(k) = \max_{0<a<\sqrt{\frac{k-2}{2(k-1)}}} \frac{\log(1-a^2)}{\log\left\{\left(1-\frac{1}{(k-1)^2}\right)\left(1-\frac{2(k-1)}{k-2}a^2\right)\right\}-\log(1-a^2)}. α ( k ) = 0 < a < 2 ( k − 1 ) k − 2 max log { ( 1 − ( k − 1 ) 2 1 ) ( 1 − k − 2 2 ( k − 1 ) a 2 ) } − log ( 1 − a 2 ) log ( 1 − a 2 ) .
Upper bounds. In search of better coloring algorithms, Karger, Motwani, and Sudan [38 ] presented the first rounding scheme on vector k k k -coloring. For a vector k k k -colorable graph with maximum degree n β n^\beta n β , their rounding scheme finds an independent set of size n 1 − k − 2 k β − o ( 1 ) n^{1-\frac{k-2}{k}\beta-o(1)} n 1 − k k − 2 β − o ( 1 ) . This shows that the chromatic number of vector k k k -colorable graphs is at most n 1 − k − 2 k + o ( 1 ) n^{1-\frac{k-2}{k}+o(1)} n 1 − k k − 2 + o ( 1 ) . Moreover, they introduced the following recursive technique that improves the bound when β \beta β is large.
Lemma 8 ([38 ] , restated). For integers k ≥ 2 k \ge2 k ≥ 2 , suppose that an independent set of size n 1 − γ k ( β ) − o ( 1 ) n^{1-\gamma_k(\beta)-o(1)} n 1 − γ k ( β ) − o ( 1 ) can be found in polynomial time for a given vector k k k -colorable graph with maximum degree ≤ n β \le n^\beta ≤ n β . Then, an independent set of size n 1 − a k − o ( 1 ) n^{1-a_k-o(1)} n 1 − a k − o ( 1 ) can be found in polynomial time for a given vector k k k -colorable graph, where
a 2 = 0 , a k = min 0 < β ≤ 1 max { γ k ( β ) , 1 − β ( 1 − a k − 1 ) } ( k ≥ 3 ) . a_2=0,\qquad a_k=\min_{0<\beta\le1}\max\{\gamma_k(\beta),1-\beta(1-a_{k-1})\}\quad(k\ge3). a 2 = 0 , a k = 0 < β ≤ 1 min max { γ k ( β ) , 1 − β ( 1 − a k − 1 )} ( k ≥ 3 ) .
We can also obtain a coloring with n a k + o ( 1 ) n^{a_k+o(1)} n a k + o ( 1 ) colors.
Proof. Let k ≥ 3 k \ge3 k ≥ 3 , and let G = ( V , E ) G=(V,E) G = ( V , E ) be a vector k k k -colorable graph. Then, for x ∈ V x\in V x ∈ V , its neighborhood G [ N ( x ) ] G[N(x)] G [ N ( x )] is vector ( k − 1 ) (k-1) ( k − 1 ) -colorable, as shown in [38 Lemma 4.3] . Then, by applying the algorithm for vector ( k − 1 ) (k-1) ( k − 1 ) -colorable graphs to G [ N ( x ) ] G[N(x)] G [ N ( x )] , we obtain an independent set of size ∣ N ( x ) ∣ 1 − a k − 1 − o ( 1 ) |N(x)|^{1-a_{k-1}-o(1)} ∣ N ( x ) ∣ 1 − a k − 1 − o ( 1 ) , which is n β ( 1 − a k − 1 ) − o ( 1 ) n^{\beta(1-a_{k-1})-o(1)} n β ( 1 − a k − 1 ) − o ( 1 ) for the largest-degree vertex x x x . Hence, Theorem 8 follows. □
This improves the bound from a k = 1 − 2 k a_k=1-\frac{2}{k} a k = 1 − k 2 to a k = 1 − 3 k a_k=1-\frac{3}{k} a k = 1 − k 3 , i.e., the chromatic number of vector k k k -colorable graphs is at most n 1 − 3 k + o ( 1 ) n^{1-\frac{3}{k}+o(1)} n 1 − k 3 + o ( 1 ) . Then, Arora, Chlamtáč, and Charikar [1 ] improved the bound to O ( n 0.2448 ) O(n^{0.2448}) O ( n 0.2448 ) for the case k = 3 k=3 k = 3 . Chlamtáč and Haviv [16 ] further improved the bound for strict vector k k k -coloring. However, the lower bound by [21 ] (empirically) has the exponent of 1 − 4 k + O ( k − 2 ) 1-\frac{4}{k}+O(k^{-2}) 1 − k 4 + O ( k − 2 ) , leaving a substantial Θ ( k − 1 ) \Theta(k^{-1}) Θ ( k − 1 ) gap, as illustrated in the graph of Figure 1 .
Figure 1. The comparison of the performance of rounding schemes ([38 ] , [1 ] , [16 ] , and our result) and a lower bound by [21 ] .
Our improvements. Our new rounding scheme for vector k k k -colorable graphs (Theorem 2 ) obtains an independent set of size n 1 − k − 2 2 + k β − o ( 1 ) n^{1-\frac{k-2}{2+k\beta}-o(1)} n 1 − 2 + k β k − 2 − o ( 1 ) for a vector k k k -colorable graph with maximum degree n β n^\beta n β . Hence, substituting β = 1 \beta=1 β = 1 , the chromatic number of vector k k k -colorable graphs is at most n 1 − 4 k + 2 + o ( 1 ) n^{1-\frac{4}{k+2}+o(1)} n 1 − k + 2 4 + o ( 1 ) . This reduces the gap to the lower bound to only ≈ 2 k − 2 \approx2k^{-2} ≈ 2 k − 2 (empirically), showing the asymptotic tightness (with respect to k k k ) of the lower-bound exponent by [21 ] . With recursion (Theorem 8 ), we obtain an even tighter upper bound: O ( n 0.1836 ) O(n^{0.1836}) O ( n 0.1836 ) colors for the case k = 3 k=3 k = 3 , which reduces the gap to only ≈ 11 3 k − 3 \approx\frac{11}{3}k^{-3} ≈ 3 11 k − 3 (empirically). See Figure 1 for the comparison of the exponents.
Organization
We begin with an overview of the algorithms and proof techniques in § 2 . Then, in § 3 , we review preliminaries on the vector coloring SDP, the geometry, and Gaussian bounds. In § 4 , we present an algorithm that finds a large independent set using only the strict vector 3-coloring SDP (Theorem 3), which yields an O ( n 1 / 6 + ε ) O(n^{1/6+\varepsilon}) O ( n 1/6 + ε ) -coloring algorithm combined with Theorem 1 . We also extend the result from strict vector 3-colorings to ordinary vector k k k -colorings in Appendix § 8 . In § 5 , we develop a strengthened SDP-based algorithm and establish a stronger independent-set guarantee (Theorem 5) that yields an O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) -coloring algorithm. Finally, Appendix § 9 discusses polynomial-time implementation and approximate SDP constraints, while Appendix § 10 provides computational verification of the numerical inequalities used in our analysis.
Overview of Proof
Here, we briefly present our two algorithms and the main ideas behind their analysis.
The n 1 / 6 + o ( 1 ) n^{1/6+o(1)} n 1/6 + o ( 1 ) -Coloring Algorithm (Section 4)
Our algorithm. By Theorem 1 , it suffices to design an algorithm that finds a sufficiently large independent set in graphs with maximum degree below n 2 / 3 n^{2/3} n 2/3 . First, we calculate a (strict) vector 3-coloring of the graph that satisfies
⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1
for all edges x y ∈ E xy \in E x y ∈ E . Then, our algorithm performs Gaussian local-maximum selection , which is a novel technique. In Gaussian local-maximum selection, we sample a random Gaussian vector g ∼ N ( 0 , I ) g \sim N(0,I) g ∼ N ( 0 , I ) , and compute the Gaussian projection ⟨ g , v x ⟩ \langle g,v_x\rangle ⟨ g , v x ⟩ for each x ∈ V x \in V x ∈ V . Next, let I I I be the set of vertices whose Gaussian projection is larger than that of any neighboring vertex (see Figure 2 ). Then I I I is always an independent set, and in many cases, ∣ I ∣ ≥ n 5 / 6 − o ( 1 ) |I| \ge n^{5/6-o(1)} ∣ I ∣ ≥ n 5/6 − o ( 1 ) holds. Intuitively, the independent set is large due to ⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1 for all x y ∈ E xy \in E x y ∈ E .
Figure 2. A sketch of the Gaussian local-maximum selection technique. The arrow inside each vertex results from the vector 3-coloring.
However, only one attempt at Gaussian local-maximum selection is not enough to guarantee an independent set of size n 5 / 6 − o ( 1 ) n^{5/6-o(1)} n 5/6 − o ( 1 ) for every graph. Therefore, our algorithm constructs a sequence of nested cores F 0 ⊇ F 1 ⊃ ⋯ F_0 \supseteq F_1 \supset\cdots F 0 ⊇ F 1 ⊃ ⋯ , with depth ℓ \ell ℓ nearly log n \log n log n . We begin with F 0 = V F_0=V F 0 = V , and at each level j j j (= 1 , 2 , … =1,2,\ldots = 1 , 2 , … ), we sample Gaussian local maxima of G [ F j − 1 ] G[F_{j-1}] G [ F j − 1 ] many times (e.g., ≈ n 1 / 3 \approx n^{1/3} ≈ n 1/3 times). If all of them fail to obtain a large independent set, we delete vertices that are frequently selected (e.g., those whose empirical estimate is at least δ 2 \frac{\delta}{2} 2 δ , where δ ≈ n − 1 / 6 \delta\approx n^{-1/6} δ ≈ n − 1/6 ) and let F j F_j F j be all remaining vertices. For each remaining vertex x x x , we can estimate that the probability of selecting x x x (by local-maximum selection) is less than δ \delta δ because of the large number of samples. We call this the probability estimate ; it will be used in the analysis.
Proof approach. We show that it is impossible for the algorithm to fail at all levels, assuming that the probability estimate is accurate. The key idea is to utilize the Gaussian cover [1 ] ; a Gaussian cover is a collection of vectors Z = { z 1 , z 2 , … } Z=\{z_1,z_2,\ldots\} Z = { z 1 , z 2 , … } such that for a random Gaussian vector g g g , the Gaussian projection onto at least one vector in Z Z Z exceeds a prescribed threshold s s s , with
high probability q ≈ 1 q \approx1 q ≈ 1 . We prove the algorithm’s guarantee by constructing Gaussian covers with progressively increasing thresholds s s s and deriving a contradiction.
Constructing a Gaussian cover. The question is how to obtain the Gaussian cover that proves contradiction. First, we pick a vertex r ∈ F ℓ r \in F_{\ell} r ∈ F ℓ from the deepest level ℓ \ell ℓ (note that F ℓ ≠ ∅ F_{\ell} \ne\varnothing F ℓ = ∅ due to Theorem 20) and obtain the initial Gaussian cover with vertex set W 0 ← N F ℓ − 1 ( r ) W_0 \leftarrow N_{F_{\ell-1}}(r) W 0 ← N F ℓ − 1 ( r ) (Theorem 24). Then, we iteratively “lift” the Gaussian cover to shallower levels, increasing the threshold at the cost of only a negligible loss of probability. Specifically, we let
W 1 ← N F ℓ − 2 ( W 0 ) , W 2 ← N F ℓ − 3 ( W 1 ) , … , W_1 \leftarrow N_{F_{\ell-2}}(W_0),\ W_2 \leftarrow N_{F_{\ell-3}}(W_1),\ldots, W 1 ← N F ℓ − 2 ( W 0 ) , W 2 ← N F ℓ − 3 ( W 1 ) , … ,
where W j W_j W j is the vertex set of the j j j -th Gaussian cover (see also Figure 3 ). Importantly, the Gaussian projections we use in Gaussian covers are not ⟨ g , v y ⟩ \langle g,v_y\rangle ⟨ g , v y ⟩ for each y ∈ W j y \in W_j y ∈ W j (j ≥ 0 j \ge0 j ≥ 0 ), but ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ , where u y u_y u y is the projection of v y v_y v y that is orthogonal to v r v_r v r . This means that we show
Figure 3. A sketch of lifting a Gaussian cover. Vertices with black outlines belong to F j F_j F j , F j − 1 F_{j-1} F j − 1 , and F j − 2 F_{j-2} F j − 2 . Vertices with gray outlines do not belong to them.
P g [ max y ∈ W j ⟨ g , u y ⟩ > s j ] = 1 − o ( 1 ) \mathbb{P}_g\left[\max_{y\in W_j}\langle g,u_y\rangle>s_j\right]=1-o(1) P g [ y ∈ W j max ⟨ g , u y ⟩ > s j ] = 1 − o ( 1 )
for some s j s_j s j . On the other hand, by the definition of g g g and the fact that $\lVert u_y\rVert \le 1,
P g [ max y ∈ W j ⟨ g , u y ⟩ > s j ] < n e − s j 2 / 2 \mathbb{P}_g\left[\max_{y\in W_j}\langle g,u_y\rangle>s_j\right]<ne^{-s_j^2/2} P g [ y ∈ W j max ⟨ g , u y ⟩ > s j ] < n e − s j 2 /2
can be easily shown. Therefore, if s j s_j s j becomes large enough before reaching F 0 F_0 F 0 , we get a contradiction, showing that the algorithm never fails.
Lifting lemma. The remaining step is to show that { u y } y ∈ W 0 , { u y } y ∈ W 1 , … \{u_y\}_{y\in W_0},\{u_y\}_{y\in W_1},\ldots { u y } y ∈ W 0 , { u y } y ∈ W 1 , … are actually Gaussian covers. The key ingredient is a lifting lemma (Theorem 25) that propagates a Gaussian cover from one core to the preceding core, increasing the threshold at the cost of losing a negligible loss of probability. To understand this propagation, fix a root r r r in the deepest core, a vertex x x x with ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , and a neighbor y y y of x x x . Then, as written in Subsection 3.2, the strict vector coloring constraint ⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1 directly implies
⟨ g , u y ⟩ = − ( 1 2 + 3 2 ξ y ⟨ v r , v x ⟩ ∥ u x ∥ ) ⟨ g , u x ⟩ + 3 2 ⟨ g , w y ⟩ . (1) \langle g,u_y\rangle=-\left(\frac{1}{2}+\frac{\sqrt{3}}{2}\frac{\xi_y\langle v_r,v_x\rangle}{\lVert u_x\rVert}\right)\langle g,u_x\rangle+\frac{\sqrt{3}}{2}\langle g,w_y\rangle. \tag*{(1)} ⟨ g , u y ⟩ = − ( 2 1 + 2 3 ∥ u x ∥ ξ y ⟨ v r , v x ⟩ ) ⟨ g , u x ⟩ + 2 3 ⟨ g , w y ⟩ . (1)
Here ξ y \xi_y ξ y denotes the signed in-plane component of the edge residual, and w y w_y w y is orthogonal to the plane spanned by v r v_r v r and v x v_x v x , satisfying ∥ w y ∥ 2 = 1 − ( ξ y ) 2 \lVert w_y\rVert^2 = 1 - (\xi_y)^2 ∥ w y ∥ 2 = 1 − ( ξ y ) 2 .
The difficulty for analysis is that the Gaussian projection ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ in Equation (1) depends on both ξ y \xi_y ξ y and 1 − ( ξ y ) 2 \sqrt{1-(\xi_y)^2} 1 − ( ξ y ) 2 . Indeed, ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ consists of a contribution inherited from x x x (the first term on the right-hand side of (1)) and an independent transverse Gaussian contribution (the second term), where the former depends on ξ y \xi_y ξ y and the latter depends on 1 − ( ξ y ) 2 \sqrt{1-(\xi_y)^2} 1 − ( ξ y ) 2 .
Resolving the difficulty. To resolve the difficulty, we made use of Ehrhard’s inequality [19 ] (Theorem 28). First, we obtain two convex sets with contrasting Gaussian measures, using the degree bound (≤ n 2 / 3 \le n^{2/3} ≤ n 2/3 ) and the small local-maximum probability for each vertex (see Theorem 30 and the proof of Theorem 26). Then, we applied Ehrhard’s inequality and yielded a lower bound on the probability that some neighbor attains a sufficiently large transverse Gaussian projection, with a threshold depending on both ξ y \xi_y ξ y and 1 − ( ξ y ) 2 \sqrt{1-(\xi_y)^2} 1 − ( ξ y ) 2 (Theorem 26).
This method is enough to show that if Gaussian projection at vertex x x x is below − s -s − s for the current cover threshold s s s , then with high probability, some neighbor y y y has a Gaussian projection above a higher threshold s ′ > s s' > s s ′ > s , by a positive margin (Theorem 27). This establishes the desired lifting lemma (Theorem 25). Moreover, this algorithm extends to ordinary vector k k k -coloring from the k = 3 k=3 k = 3 case, and we use this extension in the band rounding of the improved algorithm.
The O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) -Coloring Algorithm
Our algorithm. Our O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) -coloring algorithm performs both Gaussian local-maximum selection and a new technique called band rounding . The band rounding procedure is as follows. We first partition the range [ − 1 2 , 1 ] [-\frac{1}{2},1] [ − 2 1 , 1 ] into O ( log n ) O(\log n) O ( log n ) intervals whose length is at most 1 log n \frac{1}{\log n} l o g n 1 . For each vertex r r r and interval P P P , we define the band S r , P S_{r,P} S r , P as the set of vertices x x x satisfying ⟨ v r , v x ⟩ ∈ P \langle v_r,v_x\rangle\in P ⟨ v r , v x ⟩ ∈ P . For each band, we aim to obtain a large independent set; we omit the details due to the page limit, but we apply our generalized vector-coloring algorithm in Appendix § 8 and also use additional SDP constraints valid when G G G is 3-colorable. See Subsection 5.1–Subsection 5.3 for details.
Proof approach. The goal is to show that the probability that both Gaussian local-maximum selection and band rounding fail to obtain an independent set of size n 1 − α n^{1-\alpha} n 1 − α is very low for α ≈ 0.1536995 \alpha\approx0.1536995 α ≈ 0.1536995 . The crux of the proof is to show that if both algorithms fail, then
P [ max x ∈ N F j ( r ) ⟨ g , u x ⟩ ∥ u x ∥ > A t ] ≤ t − 4 Φ ‾ ( t / 2 ) , (2) \mathbb{P}\left[\max_{x\in N_{F_j(r)}}\frac{\langle g,u_x\rangle}{\lVert u_x\rVert}>At\right]\le t^{-4}\overline{\Phi}(t/2), \tag*{(2)} P [ x ∈ N F j ( r ) max ∥ u x ∥ ⟨ g , u x ⟩ > A t ] ≤ t − 4 Φ ( t /2 ) , (2)
holds at A = 1.968 A=1.968 A = 1.968 for sufficiently deep core j ≥ c j\ge c j ≥ c (c = o ( ℓ ) c=o(\ell) c = o ( ℓ ) ). As in Subsection 2.1, F j F_j F j is the j j j -th core of the Gaussian local-maximum selection algorithm, and u x u_x u x is the projection of v x v_x v x orthogonal to v r v_r v r , where ( v 1 , … , v n ) (v_1,\ldots,v_n) ( v 1 , … , v n ) is the result of a (strict) vector 3-coloring used in the Gaussian local-maximum selection algorithm. Importantly, this tail bound is stronger than the bound used in the n 1 / 6 + o ( 1 ) n^{1/6+o(1)} n 1/6 + o ( 1 ) -coloring algorithm.
Assuming Equation (2) holds, the lifting lemma can be significantly improved. Unfortunately, without the band rounding algorithm and Equation (2) , we cannot apply the lifting lemma when the target independent-set size exceeds Ω ( n 5 / 6 ) \Omega(n^{5/6}) Ω ( n 5/6 ) . However, thanks to Equation (2) , we can construct more favorable convex sets to apply Ehrhard’s inequality, and the lifting lemma still works even if the target independent-set size is n 1 − α n^{1-\alpha} n 1 − α . Therefore, by iteratively using the improved lifting lemma, starting from F ℓ F_\ell F ℓ , we ultimately reach a contradiction by Equation (2) before F o ( ℓ ) F_{o(\ell)} F o ( ℓ ) , which implies that the algorithm (almost) never fails. The sketch is shown in Figure 4 .
Figure 4. The sketch of our proof approach. We first prove the tail upper bound of the probability using Equation (2) , and then derive a contradiction using an improved lifting lemma.
Proving Equation (2). First, Equation (2) is equivalent to bounding the probability of a local event , where at least one neighbor of r r r has a large Gaussian projection. Formally, we define
L i ( r ) = { ∃ x ∈ N F i + 1 ( r ) : ⟨ g , u x ⟩ ∥ u x ∥ > A i t } \mathcal{L}_i(r)=\left\{\exists x\in N_{F_{i+1}}(r):\frac{\langle g,u_x\rangle}{\lVert u_x\rVert}>A_i t\right\} L i ( r ) = { ∃ x ∈ N F i + 1 ( r ) : ∥ u x ∥ ⟨ g , u x ⟩ > A i t }
and our goal is to show P [ L i ( r ) ] < ( i + 2 ) ε t \mathbb{P}[\mathcal{L}_i(r)]<(i+2)\varepsilon_t P [ L i ( r )] < ( i + 2 ) ε t at ε t ≈ e − t 2 / ( 8 − ε ) \varepsilon_t\approx e^{-t^2/(8-\varepsilon)} ε t ≈ e − t 2 / ( 8 − ε ) . Proving this at A i = 1.968 A_i=1.968 A i = 1.968 for all i ≥ c i\ge c i ≥ c , where c = o ( ℓ ) c=o(\ell) c = o ( ℓ ) , directly implies that Equation (2) holds. We show this bound by induction on i i i . We start at i = 0 i=0 i = 0 with A 0 ≈ 2.18 A_0\approx2.18 A 0 ≈ 2.18 , and use gradually decreasing A i A_i A i ’s in the induction step, eventually reaching A c = 1.968 A_c=1.968 A c = 1.968 at c = o ( ℓ ) c=o(\ell) c = o ( ℓ ) . However, it is very difficult to show P [ L i ( r ) ] < ( i + 2 ) ε t \mathbb{P}[\mathcal{L}_i(r)]<(i+2)\varepsilon_t P [ L i ( r )] < ( i + 2 ) ε t directly. Therefore, we introduce the global event (Figure 5 (b)), that at least one of the vertices in core F i + 1 F_{i+1} F i + 1 has a large Gaussian projection. Formally,
Figure 5. (a) The sketch of the local event. (b) The sketch of the global event. Vertices with black outlines belong to F i + 1 F_{i+1} F i + 1 . Vertices with gray outlines do not belong to them.
G i ( r ) = { ∃ x ∈ F i + 1 : ∣ ⟨ g , u x ⟩ ∣ > q A i ( ⟨ v r , v x ⟩ ) t } , \mathcal{G}_i(r)=\left\{\exists x\in F_{i+1}:|\langle g,u_x\rangle|>q_{A_i}(\langle v_r,v_x\rangle)t\right\}, G i ( r ) = { ∃ x ∈ F i + 1 : ∣ ⟨ g , u x ⟩ ∣ > q A i (⟨ v r , v x ⟩) t } ,
where q A i ( ρ ) : [ − 1 , 1 ] → R + q_{A_i}(\rho):[-1,1]\to\mathbb{R}^{+} q A i ( ρ ) : [ − 1 , 1 ] → R + is a function dependent on the correlation ρ \rho ρ . Now, let E i + 1 ( r ) ∈ { L i + 1 ( r ) , G i + 1 ( r ) } \mathcal{E}_{i+1}(r)\in\{\mathcal{L}_{i+1}(r),\mathcal{G}_{i+1}(r)\} E i + 1 ( r ) ∈ { L i + 1 ( r ) , G i + 1 ( r )} , which indicates any of two events. Then, obviously
P [ E i + 1 ( r ) ] ≤ P [ G i ( r ) ] + P [ E i + 1 ( r ) ∖ G i ( r ) ] \mathbb{P}[\mathcal{E}_{i+1}(r)]\le\mathbb{P}[\mathcal{G}_i(r)]+\mathbb{P}[\mathcal{E}_{i+1}(r)\setminus\mathcal{G}_i(r)] P [ E i + 1 ( r )] ≤ P [ G i ( r )] + P [ E i + 1 ( r ) ∖ G i ( r )]
holds. In the induction step on i i i , we bound the probability of E i + 1 ( r ) \mathcal{E}_{i+1}(r) E i + 1 ( r ) by showing that both the first term P [ G i ( r ) ] \mathbb{P}[\mathcal{G}_i(r)] P [ G i ( r )] and the second term P [ E i + 1 ( r ) ∖ G i ( r ) ] \mathbb{P}[\mathcal{E}_{i+1}(r)\setminus\mathcal{G}_i(r)] P [ E i + 1 ( r ) ∖ G i ( r )] are small. Especially, we aim P [ E i ( r ) ] ≤ ( i + 2 ) ε t \mathbb{P}[\mathcal{E}_i(r)]\le(i+2)\varepsilon_t P [ E i ( r )] ≤ ( i + 2 ) ε t . When E i + 1 = L i + 1 \mathcal{E}_{i+1}=\mathcal{L}_{i+1} E i + 1 = L i + 1 , we bound the probability of the desired local events.
Bounding each term. The first term can be bounded by induction on i i i . For the second term, when E i + 1 = G i + 1 \mathcal{E}_{i+1}=\mathcal{G}_{i+1} E i + 1 = G i + 1 , we use the fact that if the band rounding algorithm fails, then the size of each band is small for some correlation interval P P P (this is exactly why the band rounding is required). By Theorem 37, the band size ∣ S r , P ∣ |S_{r,P}| ∣ S r , P ∣ is less than f ( ρ ) f(\rho) f ( ρ ) for some f : ρ → R + f:\rho\to\mathbb{R}^{+} f : ρ → R + , where ρ \rho ρ is the correlation, which provides a nontrivial improvement over the trivial bound ∣ S r , P ∣ ≤ n |S_{r,P}| \le n ∣ S r , P ∣ ≤ n for certain correlation ranges. Using this fact, we bound the second term by applying Ehrhard’s inequality with Equation (1) (see Theorem 42 (i)).
When E i + 1 = L i + 1 \mathcal{E}_{i+1} = \mathcal{L}_{i+1} E i + 1 = L i + 1 , the intuitive reason for the small second term is that it is very rare to have a vertex x x x with a large Gaussian projection in the neighbors of r r r , despite no vertices y ∈ F i + 1 y \in F_{i+1} y ∈ F i + 1 that are adjacent to x x x having a large absolute value of Gaussian projection. This is strange because if ⟨ g , u x ⟩ \langle g, u_x\rangle ⟨ g , u x ⟩ is large, then ⟨ g , u y ⟩ \langle g, u_y\rangle ⟨ g , u y ⟩ is likely to be negatively large by the negative correlation between u x u_x u x and u y u_y u y derived from the vector 3-coloring. In the formal proof, we apply the Gaussian Hölder inequality and Ehrhard’s inequality; see Theorem 38 and Theorem 42 (ii).
Preliminaries
For a subset F ⊆ V F \subseteq V F ⊆ V , write N F ( x ) = { y ∈ F : x y ∈ E } N_F(x) = \{y \in F : xy \in E\} N F ( x ) = { y ∈ F : x y ∈ E } for x ∈ V x \in V x ∈ V and N F ( S ) = ⋃ x ∈ S N F ( x ) N_F(S) = \bigcup_{x \in S} N_F(x) N F ( S ) = ⋃ x ∈ S N F ( x ) for S ⊆ V S \subseteq V S ⊆ V ; we write N ( x ) = N V ( x ) N(x) = N_V(x) N ( x ) = N V ( x ) and N ( S ) = N V ( S ) N(S) = N_V(S) N ( S ) = N V ( S ) , and let G [ F ] G[F] G [ F ] be the subgraph of G G G induced by F F F , whose edge set we denote by E [ F ] = { x y ∈ E : x , y ∈ F } E[F] = \{xy \in E : x,y \in F\} E [ F ] = { x y ∈ E : x , y ∈ F } . For a graph G G G , we write Δ ( G ) \Delta(G) Δ ( G ) for its maximum degree. Fix a total order on V V V for breaking ties between Gaussian projections. Define
I F ( g ) = { x ∈ F : ∀ y ∈ N F ( x ) : ( ⟨ g , v x ⟩ , x ) > ( ⟨ g , v y ⟩ , y ) } . I_F(g) = \{x \in F : \forall y \in N_F(x) : (\langle g, v_x\rangle, x) > (\langle g, v_y\rangle, y)\}. I F ( g ) = { x ∈ F : ∀ y ∈ N F ( x ) : (⟨ g , v x ⟩ , x ) > (⟨ g , v y ⟩ , y )} .
Then I F ( g ) I_F(g) I F ( g ) is independent for every g g g .
Let n = ∣ V ∣ n = |V| n = ∣ V ∣ be the number of vertices in the input graph.
Vector Coloring and SDP Relaxation
For an integer k ≥ 2 k \ge2 k ≥ 2 , we can relax the k k k -coloring problem to the following semidefinite program.
⟨ v x , v y ⟩ = − 1 k − 1 for all x y ∈ E , ∥ v x ∥ = 1 for all x ∈ V . \begin{aligned}
\langle v_x, v_y\rangle&= -\frac{1}{k-1} && \text{for all } xy \in E,\\
\lVert v_x\rVert&= 1 && \text{for all } x \in V.
\end{aligned} ⟨ v x , v y ⟩ ∥ v x ∥ = − k − 1 1 = 1 for all x y ∈ E , for all x ∈ V .
We can easily verify this is a relaxation of the original problem by assigning to each vertex a vector corresponding to its color in a regular ( k − 1 ) (k-1) ( k − 1 ) -simplex, according to an unknown proper k k k -coloring. The constraints themselves make sense for every real number k ≥ 2 k \ge2 k ≥ 2 , and we define vector colorings in this generality.
Definition 10 (Strict vector k k k -coloring). Let G = ( V , E ) G=(V,E) G = ( V , E ) be a graph and let k ≥ 2 k\ge2 k ≥ 2 be a real number. A set of unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V is called a strict vector k k k -coloring if it satisfies ⟨ v x , v y ⟩ = − 1 k − 1 \langle v_x,v_y\rangle=-\frac{1}{k-1} ⟨ v x , v y ⟩ = − k − 1 1 for every edge x y ∈ E xy\in E x y ∈ E .
Definition 11 (Ordinary vector k k k -coloring). Ordinary vector k k k -coloring is the variant obtained by changing the equality in the definition of a strict vector k k k -coloring to ≤ \le ≤ . The smallest real k ≥ 2 k\ge2 k ≥ 2 for which G G G has an ordinary vector k k k -coloring is the vector chromatic number of G G G introduced by Karger, Motwani, and Sudan [38 ] , which is at most the chromatic number of G G G .
Unless stated otherwise, a vector k k k -coloring means an ordinary vector k k k -coloring.
An SDP can be solved in polynomial time only up to an additive error. Accordingly, when computing a strict vector k k k -coloring, we allow an additive error ε S D P \varepsilon_{\mathrm{SDP}} ε SDP in the edge constraints. Throughout, we fix ε S D P = e − 3 n \varepsilon_{\mathrm{SDP}}=e^{-3n} ε SDP = e − 3 n , so that log ( 1 / ε S D P ) = 3 n \log(1/\varepsilon_{\mathrm{SDP}})=3n log ( 1/ ε SDP ) = 3 n , and the computed vectors satisfy − 1 k − 1 ≤ ⟨ v x , v y ⟩ ≤ − 1 k − 1 + ε S D P -\frac{1}{k-1}\le\langle v_x,v_y\rangle\le-\frac{1}{k-1}+\varepsilon_{\mathrm{SDP}} − k − 1 1 ≤ ⟨ v x , v y ⟩ ≤ − k − 1 1 + ε SDP for every edge x y ∈ E xy\in E x y ∈ E . In Appendix § 9 .§ 2 , we show that treating this constraint as exact is harmless for our analysis, which justifies the discussion below.
Vector Identities
Here, we introduce useful vector notation and deduce key vector identities.
Given unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V , we often project some v y v_y v y along another v x v_x v x and an orthogonal direction. For x , y ∈ V x,y\in V x , y ∈ V with v y ≠ ± v x v_y\ne\pm v_x v y = ± v x , we write
v x y = v y − ⟨ v x , v y ⟩ v x ∥ v y − ⟨ v x , v y ⟩ v x ∥ v_{xy}=\frac{v_y-\langle v_x,v_y\rangle v_x}{\lVert v_y-\langle v_x,v_y\rangle v_x\rVert} v x y = ∥ v y − ⟨ v x , v y ⟩ v x ∥ v y − ⟨ v x , v y ⟩ v x
The vector v x y v_{xy} v x y is a unit vector orthogonal to v x v_x v x , and v y = ⟨ v x , v y ⟩ v x + 1 − ⟨ v x , v y ⟩ 2 v x y v_y=\langle v_x,v_y\rangle v_x+\sqrt{1-\langle v_x,v_y\rangle^2}v_{xy} v y = ⟨ v x , v y ⟩ v x + 1 − ⟨ v x , v y ⟩ 2 v x y . For x y ∈ E xy\in E x y ∈ E , we call v x y v_{xy} v x y the edge residual of y y y at x x x . Note that v x y ≠ v y x v_{xy}\ne v_{yx} v x y = v y x in general. If { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V is a strict vector 3 3 3 -coloring and x y ∈ E xy\in E x y ∈ E , then ⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1 , so v x y = 2 v y + v x 3 v_{xy}=\frac{2v_y+v_x}{\sqrt{3}} v x y = 3 2 v y + v x and
v y = − 1 2 v x + 3 2 v x y . (3) v_y=-\frac{1}{2}v_x+\frac{\sqrt{3}}{2}v_{xy}. \tag*{(3)} v y = − 2 1 v x + 2 3 v x y . (3)
In the rest of this subsection, let { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V be a strict vector 3 3 3 -coloring of G G G , and fix a vertex r ∈ V r\in V r ∈ V . For x ∈ V x\in V x ∈ V , write ρ x = ⟨ v r , v x ⟩ \rho_x=\langle v_r,v_x\rangle ρ x = ⟨ v r , v x ⟩ and
u x = v x − ρ x v r , u_x=v_x-\rho_xv_r, u x = v x − ρ x v r ,
so that u x ⊥ v r u_x\perp v_r u x ⊥ v r and ∥ u x ∥ = 1 − ρ x 2 \lVert u_x\rVert=\sqrt{1-\rho_x^2} ∥ u x ∥ = 1 − ρ x 2 ; thus, in contrast to v x v_x v x , u x u_x u x is not necessarily a unit vector. If ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , then u x = ∥ u x ∥ v r x u_x=\lVert u_x\rVert v_{rx} u x = ∥ u x ∥ v r x and v x r = ( v r − ρ x v x ) / ∥ u x ∥ v_{xr}=(v_r-\rho_xv_x)/\lVert u_x\rVert v x r = ( v r − ρ x v x ) / ∥ u x ∥ by definition, where v r x ⊥ v r v_{rx}\perp v_r v r x ⊥ v r and v x r ⊥ v x v_{xr}\perp v_x v x r ⊥ v x are unit vectors, and v x v_x v x and v x r v_{xr} v x r form an orthonormal basis of span ( v r , v x ) \operatorname{span}(v_r,v_x) span ( v r , v x ) . Let x ∈ V x\in V x ∈ V satisfy ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , and let y ∈ N V ( x ) y\in N_V(x) y ∈ N V ( x ) . Since v x y ⊥ v x v_{xy}\perp v_x v x y ⊥ v x , we may decompose
v x y = ξ y v x r + w y , (4) v_{xy}=\xi_yv_{xr}+w_y, \tag*{(4)} v x y = ξ y v x r + w y , (4)
where ξ y = ⟨ v x y , v x r ⟩ \xi_y=\langle v_{xy},v_{xr}\rangle ξ y = ⟨ v x y , v x r ⟩ , w y ⊥ span ( v r , v x ) w_y\perp\operatorname{span}(v_r,v_x) w y ⊥ span ( v r , v x ) and ( ξ y ) 2 + ∥ w y ∥ 2 = 1 (\xi_y)^2+\lVert w_y\rVert^2=1 ( ξ y ) 2 + ∥ w y ∥ 2 = 1 . Since ⟨ v r , v x r ⟩ = ( 1 − ρ x 2 ) / ∥ u x ∥ = ∥ u x ∥ \langle v_r,v_{xr}\rangle=(1-\rho_x^2)/\lVert u_x\rVert=\lVert u_x\rVert ⟨ v r , v x r ⟩ = ( 1 − ρ x 2 ) / ∥ u x ∥ = ∥ u x ∥ , taking the inner product of Equation (3) with v r v_r v r gives ρ y = − 1 2 ρ x + 3 2 ξ y ∥ u x ∥ \rho_y=-\frac{1}{2}\rho_x+\frac{\sqrt{3}}{2}\xi_y\lVert u_x\rVert ρ y = − 2 1 ρ x + 2 3 ξ y ∥ u x ∥ , that is,
3 2 ξ y = ρ y + ρ x / 2 1 − ρ x 2 . (5) \frac{\sqrt{3}}{2}\xi_y=\frac{\rho_y+\rho_x/2}{\sqrt{1-\rho_x^2}}. \tag*{(5)} 2 3 ξ y = 1 − ρ x 2 ρ y + ρ x /2 . (5)
The projection of v x r v_{xr} v x r onto v r ⊥ v_r^\perp v r ⊥ is v x r − ∥ u x ∥ v r = − ρ x u x / ∥ u x ∥ v_{xr}-\lVert u_x\rVert v_r=-\rho_xu_x/\lVert u_x\rVert v x r − ∥ u x ∥ v r = − ρ x u x / ∥ u x ∥ , and w y ⊥ v r w_y\perp v_r w y ⊥ v r . Hence projecting Equation (3) onto v r ⊥ v_r^\perp v r ⊥ gives the first equality of
u y = − ( 1 2 + 3 2 ⋅ ξ y ρ x ∥ u x ∥ ) u x + 3 2 w y = − 1 / 2 + ρ x ρ y 1 − ρ x 2 u x + 3 2 w y , (6) u_y=-\left(\frac{1}{2}+\frac{\sqrt{3}}{2}\cdot\frac{\xi_y\rho_x}{\lVert u_x\rVert}\right)u_x+\frac{\sqrt{3}}{2}w_y=-\frac{1/2+\rho_x\rho_y}{1-\rho_x^2}u_x+\frac{\sqrt{3}}{2}w_y, \tag*{(6)} u y = − ( 2 1 + 2 3 ⋅ ∥ u x ∥ ξ y ρ x ) u x + 2 3 w y = − 1 − ρ x 2 1/2 + ρ x ρ y u x + 2 3 w y , (6)
and the second equality follows from Equation (5) and ∥ u x ∥ 2 = 1 − ρ x 2 \lVert u_x\rVert^2=1-\rho_x^2 ∥ u x ∥ 2 = 1 − ρ x 2 . In Equation (6) , u x ∈ span ( v r , v x ) u_x\in\operatorname{span}(v_r,v_x) u x ∈ span ( v r , v x ) is orthogonal to v r v_r v r , and w y ∈ span ( v r , v x ) ⊥ w_y\in\operatorname{span}(v_r,v_x)^\perp w y ∈ span ( v r , v x ) ⊥ .
Gaussian Notation and Bounds
Throughout, log \log log denotes the natural logarithm. We write ϕ \phi ϕ , Φ \Phi Φ , and Φ ‾ = 1 − Φ \overline{\Phi}=1-\Phi Φ = 1 − Φ for the density, the distribution function, and the upper tail of the standard normal distribution. The vector g ∼ N ( 0 , I ) g\sim N(0,I) g ∼ N ( 0 , I ) is a standard Gaussian vector in the ambient space of the vector coloring, and P g \mathbb{P}_g P g , E g \mathbb{E}_g E g denote probability and expectation with respect to g g g ; we also write γ ( K ) = P g [ g ∈ K ] \gamma(K)=\mathbb{P}_g[g\in K] γ ( K ) = P g [ g ∈ K ] for the standard Gaussian measure of a set K K K . We use the following two tail bounds.
Fact 12. For every u ≥ 0 u\ge0 u ≥ 0 ,
c 0 u + 1 e − u 2 / 2 ≤ Φ ‾ ( u ) ≤ e − u 2 / 2 , \frac{c_0}{u+1}e^{-u^2/2}\le\overline{\Phi}(u)\le e^{-u^2/2}, u + 1 c 0 e − u 2 /2 ≤ Φ ( u ) ≤ e − u 2 /2 , where c 0 > 0 c_0>0 c 0 > 0 is an absolute constant.
We also use the following monotonicity property of the Gaussian tail.
Fact 13. The function f ( u ) = e u 2 / 2 Φ ‾ ( u ) f(u)=e^{u^2/2}\overline{\Phi}(u) f ( u ) = e u 2 /2 Φ ( u ) is strictly decreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . Consequently, for all 0 ≤ u ≤ v 0\le u\le v 0 ≤ u ≤ v ,
e ( v 2 − u 2 ) / 2 Φ ‾ ( v ) ≤ Φ ‾ ( u ) . e^{(v^2-u^2)/2}\overline{\Phi}(v)\le\overline{\Phi}(u). e ( v 2 − u 2 ) /2 Φ ( v ) ≤ Φ ( u ) . Proof. We have f ′ ( u ) = e u 2 / 2 ( u Φ ‾ ( u ) − ϕ ( u ) ) f'(u)=e^{u^2/2}(u\overline{\Phi}(u)-\phi(u)) f ′ ( u ) = e u 2 /2 ( u Φ ( u ) − ϕ ( u )) , and for every $u\ge 0,
u Φ ‾ ( u ) = u ∫ u ∞ ϕ ( s ) d s < ∫ u ∞ s ϕ ( s ) d s = ∫ u ∞ − ϕ ′ ( s ) d s = ϕ ( u ) , u\overline{\Phi}(u)=u\int_u^\infty\phi(s)\,\mathrm{d}s<\int_u^\infty s\phi(s)\,\mathrm{d}s=\int_u^\infty-\phi'(s)\,\mathrm{d}s=\phi(u), u Φ ( u ) = u ∫ u ∞ ϕ ( s ) d s < ∫ u ∞ s ϕ ( s ) d s = ∫ u ∞ − ϕ ′ ( s ) d s = ϕ ( u ) , so f ′ ( u ) < 0 f'(u)<0 f ′ ( u ) < 0 . The displayed inequality is f ( v ) ≤ f ( u ) f(v)\le f(u) f ( v ) ≤ f ( u ) rearranged.
Fact 14. If Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) , then for every u ∈ R u\in\mathbb{R} u ∈ R and η > 0 \eta>0 η > 0 ,
P [ Z ∈ [ u , u + η ] ] ≥ η ϕ ( ∣ u ∣ + η ) . \mathbb{P}[Z\in[u,u+\eta]]\ge\eta\phi(|u|+\eta). P [ Z ∈ [ u , u + η ]] ≥ η ϕ ( ∣ u ∣ + η ) . Proof. Since ∣ z ∣ ≤ ∣ u ∣ + η |z|\le|u|+\eta ∣ z ∣ ≤ ∣ u ∣ + η for every z ∈ [ u , u + η ] z\in[u,u+\eta] z ∈ [ u , u + η ] and ϕ \phi ϕ is decreasing as a function of ∣ z ∣ |z| ∣ z ∣ ,
P [ Z ∈ [ u , u + η ] ] = Φ ( u + η ) − Φ ( u ) = ∫ u u + η ϕ ( z ) d z ≥ ∫ u u + η ϕ ( ∣ u ∣ + η ) d z = η ϕ ( ∣ u ∣ + η ) . \mathbb{P}[Z\in[u,u+\eta]]=\Phi(u+\eta)-\Phi(u)=\int_u^{u+\eta}\phi(z)\,\mathrm{d}z\ge\int_u^{u+\eta}\phi(|u|+\eta)\,\mathrm{d}z=\eta\phi(|u|+\eta). P [ Z ∈ [ u , u + η ]] = Φ ( u + η ) − Φ ( u ) = ∫ u u + η ϕ ( z ) d z ≥ ∫ u u + η ϕ ( ∣ u ∣ + η ) d z = η ϕ ( ∣ u ∣ + η ) .
We also use the following Gaussian Hölder inequality.
Theorem 15 (Gaussian Hölder inequality [14 Theorem 1(i)] , two-block case). Let ( X 1 , X 2 ) (X_1,X_2) ( X 1 , X 2 ) be a centered Gaussian random vector in R n 1 × R n 2 \mathbb{R}^{n_1}\times\mathbb{R}^{n_2} R n 1 × R n 2 with covariance matrix T = ( T a b ) a , b ∈ { 1 , 2 } T=(T_{ab})_{a,b\in\{1,2\}} T = ( T ab ) a , b ∈ { 1 , 2 } , let p 1 , p 2 ≥ 1 p_1,p_2\ge1 p 1 , p 2 ≥ 1 , and let P = diag ( p 1 T 11 , p 2 T 22 ) P=\operatorname{diag}(p_1T_{11},p_2T_{22}) P = diag ( p 1 T 11 , p 2 T 22 ) . If P − T ⪰ 0 P-T\succeq0 P − T ⪰ 0 , then E [ f 1 ( X 1 ) f 2 ( X 2 ) ] ≤ ( E [ f 1 ( X 1 ) p 1 ] ) 1 / p 1 ( E [ f 2 ( X 2 ) p 2 ] ) 1 / p 2 \mathbb{E}[f_1(X_1)f_2(X_2)]\le(\mathbb{E}[f_1(X_1)^{p_1}])^{1/p_1}(\mathbb{E}[f_2(X_2)^{p_2}])^{1/p_2} E [ f 1 ( X 1 ) f 2 ( X 2 )] ≤ ( E [ f 1 ( X 1 ) p 1 ] ) 1/ p 1 ( E [ f 2 ( X 2 ) p 2 ] ) 1/ p 2 for all nonnegative measurable functions f 1 f_1 f 1 on R n 1 \mathbb{R}^{n_1} R n 1 and f 2 f_2 f 2 on R n 2 \mathbb{R}^{n_2} R n 2 .
Corollary 16. Let U 1 \mathcal{U}_1 U 1 and U 2 \mathcal{U}_2 U 2 be linear subspaces with ∣ ⟨ u 1 , u 2 ⟩ ∣ ≤ 1 2 ∥ u 1 ∥ ∥ u 2 ∥ |\langle u_1,u_2\rangle|\le\frac{1}{2}\lVert u_1\rVert\lVert u_2\rVert ∣ ⟨ u 1 , u 2 ⟩ ∣ ≤ 2 1 ∥ u 1 ∥ ∥ u 2 ∥ for all u 1 ∈ U 1 u_1\in\mathcal{U}_1 u 1 ∈ U 1 and u 2 ∈ U 2 u_2\in\mathcal{U}_2 u 2 ∈ U 2 , and let p 1 , p 2 > 1 p_1,p_2>1 p 1 , p 2 > 1 satisfy ( p 1 − 1 ) ( p 2 − 1 ) ≥ 1 4 (p_1-1)(p_2-1)\ge\frac{1}{4} ( p 1 − 1 ) ( p 2 − 1 ) ≥ 4 1 . Then P g [ A 1 ∩ A 2 ] ≤ P g [ A 1 ] 1 / p 1 P g [ A 2 ] 1 / p 2 \mathbb{P}_{g}[\mathcal{A}_1\cap\mathcal{A}_2]\le\mathbb{P}_{g}[\mathcal{A}_1]^{1/p_1}\mathbb{P}_{g}[\mathcal{A}_2]^{1/p_2} P g [ A 1 ∩ A 2 ] ≤ P g [ A 1 ] 1/ p 1 P g [ A 2 ] 1/ p 2 for all events A 1 \mathcal{A}_1 A 1 and A 2 \mathcal{A}_2 A 2 depending only on the Gaussian projections onto U 1 \mathcal{U}_1 U 1 and U 2 \mathcal{U}_2 U 2 , respectively.
Proof. Let B 1 B_1 B 1 and B 2 B_2 B 2 be matrices whose columns form orthonormal bases of U 1 \mathcal{U}_1 U 1 and U 2 \mathcal{U}_2 U 2 , respectively, and put X 1 = B 1 T g ∈ R n 1 X_1=B_1^{\mathsf{T}}g\in\mathbb{R}^{n_1} X 1 = B 1 T g ∈ R n 1 and X 2 = B 2 T g ∈ R n 2 X_2=B_2^{\mathsf{T}}g\in\mathbb{R}^{n_2} X 2 = B 2 T g ∈ R n 2 , where n 1 = dim U 1 n_1=\dim\mathcal{U}_1 n 1 = dim U 1 and n 2 = dim U 2 n_2=\dim\mathcal{U}_2 n 2 = dim U 2 . Then A 1 = { X 1 ∈ A 1 } \mathcal{A}_1=\{X_1\in A_1\} A 1 = { X 1 ∈ A 1 } and A 2 = { X 2 ∈ A 2 } \mathcal{A}_2=\{X_2\in A_2\} A 2 = { X 2 ∈ A 2 } for some measurable sets A 1 A_1 A 1 and A 2 A_2 A 2 , and the covariance matrix of ( X 1 , X 2 ) (X_1,X_2) ( X 1 , X 2 ) has blocks T 11 = I T_{11}=I T 11 = I , T 22 = I T_{22}=I T 22 = I and T 12 = B 1 T B 2 T_{12}=B_1^{\mathsf{T}}B_2 T 12 = B 1 T B 2 . For α 1 ∈ R n 1 \alpha_1\in\mathbb{R}^{n_1} α 1 ∈ R n 1 and α 2 ∈ R n 2 \alpha_2\in\mathbb{R}^{n_2} α 2 ∈ R n 2 , the assumption on U 1 \mathcal{U}_1 U 1 and U 2 \mathcal{U}_2 U 2 gives ∣ α 1 T T 12 α 2 ∣ = ∣ ⟨ B 1 α 1 , B 2 α 2 ⟩ ∣ ≤ 1 2 ∥ B 1 α 1 ∥ ∥ B 2 α 2 ∥ = 1 2 ∥ α 1 ∥ ∥ α 2 ∥ |\alpha_1^{\mathsf{T}}T_{12}\alpha_2|=|\langle B_1\alpha_1,B_2\alpha_2\rangle|\le\frac{1}{2}\lVert B_1\alpha_1\rVert\lVert B_2\alpha_2\rVert=\frac{1}{2}\lVert\alpha_1\rVert\lVert\alpha_2\rVert ∣ α 1 T T 12 α 2 ∣ = ∣ ⟨ B 1 α 1 , B 2 α 2 ⟩ ∣ ≤ 2 1 ∥ B 1 α 1 ∥ ∥ B 2 α 2 ∥ = 2 1 ∥ α 1 ∥ ∥ α 2 ∥ , and hence
( α 1 , α 2 ) T ( P − T ) ( α 1 , α 2 ) = ( p 1 − 1 ) ∥ α 1 ∥ 2 − 2 α 1 T T 12 α 2 + ( p 2 − 1 ) ∥ α 2 ∥ 2 (\alpha_1,\alpha_2)^{\mathsf{T}}(P-T)(\alpha_1,\alpha_2)=(p_1-1)\lVert\alpha_1\rVert^2-2\alpha_1^{\mathsf{T}}T_{12}\alpha_2+(p_2-1)\lVert\alpha_2\rVert^2 ( α 1 , α 2 ) T ( P − T ) ( α 1 , α 2 ) = ( p 1 − 1 ) ∥ α 1 ∥ 2 − 2 α 1 T T 12 α 2 + ( p 2 − 1 ) ∥ α 2 ∥ 2 ≥ ( p 1 − 1 ) ∥ α 1 ∥ 2 − ∥ α 1 ∥ ∥ α 2 ∥ + ( p 2 − 1 ) ∥ α 2 ∥ 2 \ge(p_1-1)\lVert\alpha_1\rVert^2-\lVert\alpha_1\rVert\lVert\alpha_2\rVert+(p_2-1)\lVert\alpha_2\rVert^2 ≥ ( p 1 − 1 ) ∥ α 1 ∥ 2 − ∥ α 1 ∥ ∥ α 2 ∥ + ( p 2 − 1 ) ∥ α 2 ∥ 2 ≥ ( 2 ( p 1 − 1 ) ( p 2 − 1 ) − 1 ) ∥ α 1 ∥ ∥ α 2 ∥ ( ∵ AM–GM inequality ) \ge\left(2\sqrt{(p_1-1)(p_2-1)}-1\right)\lVert\alpha_1\rVert\lVert\alpha_2\rVert\qquad(\because\text{ AM--GM inequality}) ≥ ( 2 ( p 1 − 1 ) ( p 2 − 1 ) − 1 ) ∥ α 1 ∥ ∥ α 2 ∥ ( ∵ AM–GM inequality ) ≥ 0. ( ∵ ( p 1 − 1 ) ( p 2 − 1 ) ≥ 1 4 ) \ge0. \qquad(\because(p_1-1)(p_2-1)\ge\tfrac{1}{4}) ≥ 0. ( ∵ ( p 1 − 1 ) ( p 2 − 1 ) ≥ 4 1 ) Thus P − T ⪰ 0 P-T\succeq0 P − T ⪰ 0 . Applying Theorem 15 to f 1 = 1 A 1 f_1=\mathbf{1}_{A_1} f 1 = 1 A 1 and f 2 = 1 A 2 f_2=\mathbf{1}_{A_2} f 2 = 1 A 2 , for which E [ f 1 ( X 1 ) p 1 ] = P g [ A 1 ] \mathbb{E}[f_1(X_1)^{p_1}]=\mathbb{P}_{g}[\mathcal{A}_1] E [ f 1 ( X 1 ) p 1 ] = P g [ A 1 ] and E [ f 2 ( X 2 ) p 2 ] = P g [ A 2 ] \mathbb{E}[f_2(X_2)^{p_2}]=\mathbb{P}_{g}[\mathcal{A}_2] E [ f 2 ( X 2 ) p 2 ] = P g [ A 2 ] , gives the claim. □
Finding a Large Independent Set from Vector 3-Coloring
Table 2 lists the parameters used in this section.
Table 2. Parameters used in Section 4.
In the low-degree regime, we write p ∗ = 1 2 Φ ‾ ( 2 3 log ( 2 n β ) ) p_* = \frac{1}{2}\overline{\Phi}\left(\sqrt{\frac{2}{3}\log(2n^\beta)}\right) p ∗ = 2 1 Φ ( 3 2 log ( 2 n β ) ) . In the high-degree regime, we write f ( β ) = 1 / 3 β − 1 f(\beta) = 1/\sqrt{3\beta- 1} f ( β ) = 1/ 3 β − 1 , which satisfies f ( β ) ≥ 1 / 2 f(\beta) \ge1/\sqrt{2} f ( β ) ≥ 1/ 2 for 1 / 3 < β ≤ 1 1/3 < \beta\le1 1/3 < β ≤ 1 . We write U 0 = 3 / ( 2 2 ) U_0 = 3/(2\sqrt{2}) U 0 = 3/ ( 2 2 ) for the initial Gaussian cover threshold divided by t t t .
Algorithmic Framework and Nested Cores
In this subsection, we present the algorithm that obtains a large independent set with constant probability, based on Gaussian local-maximum selection. Algorithm 1 always returns an independent set, since no two adjacent vertices can both be local maxima.
Fix a degree exponent β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] , and assume that every vertex of the input graph has degree at most n β n^\beta n β . Besides ℓ \ell ℓ , Algorithm 2 takes three parameters: a deletion threshold δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) , a target size T T T and a sample count R R R .
Algorithm 1 GaussianLocalMaxima
Require: vertex set $F$ and unit vectors $\{v_x\}_{x\in F}$
Ensure: an independent set $I_F$
1: Draw $g \sim N(0,I)$
2: $I_F \leftarrow \varnothing$
3: for $x \in F$ do
4: if $(\langle g,v_x\rangle,x) > (\langle g,v_y\rangle,y)$ for every $y \in N_F(x)$ then
5: $I_F \leftarrow I_F \cup \{x\}$
6: return $I_F$
Algorithm 2 performs one trial. First, set the initial core to be F 0 = V F_0 = V F 0 = V . At level j j j , we run Algorithm 1 R R R times on F j − 1 F_{j-1} F j − 1 and compute, for each x ∈ F j − 1 x \in F_{j-1} x ∈ F j − 1 , the empirical estimate p ^ j ( x ) \hat{p}_j(x) p ^ j ( x ) of the probability that x x x is selected. If any returned independent set has size at least the target size T T T , we return it with SUCCESS. Otherwise, we delete the vertices whose empirical estimate is at least δ / 2 \delta/2 δ /2 , and the remaining vertices form the next core F j F_j F j . If all levels j = 1 , … , ℓ j = 1,\ldots,\ell j = 1 , … , ℓ are exhausted, the trial exits with FAIL and is retried.
Below, we prove that if the estimates of the local-maximum probabilities are accurate, then Algorithm 2 returns a sufficiently large independent set.
Algorithm 2 IterativeGaussianLocalMaxima
Require: graph G = ( V , E ) G=(V,E) G = ( V , E ) with n = ∣ V ∣ n=|V| n = ∣ V ∣ , unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V , and parameters δ ∈ ( 0 , 1 ) \delta\in(0,1) δ ∈ ( 0 , 1 ) and T , R ∈ N T,R\in\mathbb{N} T , R ∈ N
Ensure: SUCCESS ( I ) \operatorname{SUCCESS}(I) SUCCESS ( I ) with I I I independent and ∣ I ∣ ≥ T |I|\ge T ∣ I ∣ ≥ T , or FAIL \operatorname{FAIL} FAIL
F 0 ← V F_0\leftarrow V F 0 ← V
for j = 1 j=1 j = 1 to ℓ \ell ℓ do
for i = 1 i=1 i = 1 to R R R do
I ( i ) ← GaussianLocalMaxima ( F j − 1 , { v x } x ∈ F j − 1 ) I^{(i)}\leftarrow\operatorname{GaussianLocalMaxima}(F_{j-1},\{v_x\}_{x\in F_{j-1}}) I ( i ) ← GaussianLocalMaxima ( F j − 1 , { v x } x ∈ F j − 1 )
if ∣ I ( i ) ∣ ≥ T |I^{(i)}|\ge T ∣ I ( i ) ∣ ≥ T then
return SUCCESS ( I ( i ) ) \operatorname{SUCCESS}(I^{(i)}) SUCCESS ( I ( i ) )
for x ∈ F j − 1 x\in F_{j-1} x ∈ F j − 1 do
p ^ j ( x ) ← R − 1 ∑ i = 1 R 1 [ x ∈ I ( i ) ] \widehat{p}_j(x)\leftarrow R^{-1}\sum_{i=1}^{R}\mathbf{1}[x\in I^{(i)}] p j ( x ) ← R − 1 ∑ i = 1 R 1 [ x ∈ I ( i ) ]
F j ← { x ∈ F j − 1 : p ^ j ( x ) < δ / 2 } F_j\leftarrow\{x\in F_{j-1}:\widehat{p}_j(x)<\delta/2\} F j ← { x ∈ F j − 1 : p j ( x ) < δ /2 }
return FAIL \operatorname{FAIL} FAIL
The refinement step is designed so that, if no sufficiently large independent set is found, only a small number of vertices are removed at each level, while the surviving vertices have small local-maximum probabilities with respect to the preceding core. These properties will be established in Subsection 4.3 and form the basis of our Gaussian cover argument.
Local-Maximum Probabilities
This subsection collects the results about the local-maximum probabilities of Algorithm 1 that are shared by both the low- and high-degree regimes, and then settles the low-degree regime 0 < β ≤ 1 / 3 0<\beta\le1/3 0 < β ≤ 1/3 .
A Lower Bound via Neighbor Gaussian Projections
For a remaining subset F ⊆ V F\subseteq V F ⊆ V and x ∈ F x\in F x ∈ F , define p F ( x ) = P g [ x ∈ I F ( g ) ] p_F(x)=\mathbb{P}_g[x\in I_F(g)] p F ( x ) = P g [ x ∈ I F ( g )] , the local-maximum probability of x x x .
For F ⊆ V F\subseteq V F ⊆ V and x ∈ V x\in V x ∈ V , let M F ( x ) = max y ∈ N F ( x ) ⟨ g , v x y ⟩ M_F(x)=\max_{y\in N_F(x)}\langle g,v_{xy}\rangle M F ( x ) = max y ∈ N F ( x ) ⟨ g , v x y ⟩ , where g ∼ N ( 0 , I ) g\sim N(0,I) g ∼ N ( 0 , I ) . When N F ( x ) N_F(x) N F ( x ) is empty, we define M F ( x ) = − ∞ M_F(x)=-\infty M F ( x ) = − ∞ . We call M F ( x ) M_F(x) M F ( x ) the neighborhood maximum of x x x in F F F .
Lemma 17. Let F ⊆ V F\subseteq V F ⊆ V and x ∈ F x\in F x ∈ F . Then, for every a ∈ R a\in\mathbb{R} a ∈ R ,
p F ( x ) = E g [ Φ ‾ ( M F ( x ) / 3 ) ] , p_F(x)=\mathbb{E}_g\left[\overline{\Phi}\left(M_F(x)/\sqrt{3}\right)\right], p F ( x ) = E g [ Φ ( M F ( x ) / 3 ) ] , P g [ M F ( x ) ≤ a ] ≤ p F ( x ) Φ ‾ ( a / 3 ) . \mathbb{P}_g[M_F(x)\le a]\le\frac{p_F(x)}{\overline{\Phi}(a/\sqrt{3})}. P g [ M F ( x ) ≤ a ] ≤ Φ ( a / 3 ) p F ( x ) .
Proof. Write g = Z v x + g ′ g=Zv_x+g' g = Z v x + g ′ with Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) and g ′ ∼ N ( 0 , I − v x v x T ) g'\sim N(0,I-v_xv_x^{\mathsf T}) g ′ ∼ N ( 0 , I − v x v x T ) independent. The vertex x x x is selected as a local maximum exactly when ⟨ g , v x ⟩ > ⟨ g , v y ⟩ \langle g,v_x\rangle>\langle g,v_y\rangle ⟨ g , v x ⟩ > ⟨ g , v y ⟩ for all y ∈ N F ( x ) y\in N_F(x) y ∈ N F ( x ) . Since g ′ ⊥ v x g'\perp v_x g ′ ⊥ v x and v x y ⊥ v x v_{xy}\perp v_x v x y ⊥ v x , this condition is equivalent to 3 Z > max y ∈ N F ( x ) ⟨ g ′ , v x y ⟩ = M F ( x ) \sqrt{3}Z>\max_{y\in N_F(x)}\langle g',v_{xy}\rangle=M_F(x) 3 Z > max y ∈ N F ( x ) ⟨ g ′ , v x y ⟩ = M F ( x ) . Thus, the probability that x x x is selected as a local maximum is E g [ Φ ‾ ( M F ( x ) / 3 ) ] \mathbb{E}_{g}[\overline{\Phi}(M_F(x)/\sqrt{3})] E g [ Φ ( M F ( x ) / 3 )] . On the event M F ( x ) ≤ a M_F(x) \le a M F ( x ) ≤ a , we have Φ ‾ ( M F ( x ) / 3 ) ≥ Φ ‾ ( a / 3 ) \overline{\Phi}(M_F(x)/\sqrt{3}) \ge\overline{\Phi}(a/\sqrt{3}) Φ ( M F ( x ) / 3 ) ≥ Φ ( a / 3 ) . Therefore,
p F ( x ) = E g [ Φ ‾ ( M F ( x ) / 3 ) ] ≥ E g [ Φ ‾ ( a / 3 ) 1 [ M F ( x ) ≤ a ] ] = Φ ‾ ( a / 3 ) P g [ M F ( x ) ≤ a ] , p_F(x) = \mathbb{E}_{g}[\overline{\Phi}(M_F(x)/\sqrt{3})]
\ge\mathbb{E}_{g}[\overline{\Phi}(a/\sqrt{3})\mathbf{1}[M_F(x) \le a]]
= \overline{\Phi}(a/\sqrt{3})\mathbb{P}_{g}[M_F(x) \le a], p F ( x ) = E g [ Φ ( M F ( x ) / 3 )] ≥ E g [ Φ ( a / 3 ) 1 [ M F ( x ) ≤ a ]] = Φ ( a / 3 ) P g [ M F ( x ) ≤ a ] ,
which proves the second claim. □
The Low-Degree Regime
In this subsubsection, we treat the case 0 < β ≤ 1 / 3 0 < \beta\le1/3 0 < β ≤ 1/3 . Here Algorithm 2 already succeeds at the first level, and the argument uses only p F p_F p F , Theorem 17 and the tail bounds of Theorem 12.
Lemma 18 (Low-degree regime). Fix 0 < β ≤ 1 / 3 0 < \beta\le1/3 0 < β ≤ 1/3 , and let G = ( V , E ) G = (V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β , given together with a strict vector 3-coloring. Then there is a randomized polynomial-time algorithm that, with probability at least 3 / 4 3/4 3/4 , returns an independent set I I I with ∣ I ∣ ≥ n 1 − β / 3 − o ( 1 ) |I| \ge n^{1-\beta/3-o(1)} ∣ I ∣ ≥ n 1 − β /3 − o ( 1 ) .
Proof. Run one trial of Algorithm 2 with the low-degree parameters of Table 2 . We show that this trial runs in polynomial time and, with probability at least 3 / 4 3/4 3/4 , returns SUCCESS ( I ) \operatorname{SUCCESS}(I) SUCCESS ( I ) for an independent set I I I with ∣ I ∣ ≥ T ≥ n 1 − β / 3 − o ( 1 ) |I| \ge T \ge n^{1-\beta/3-o(1)} ∣ I ∣ ≥ T ≥ n 1 − β /3 − o ( 1 ) . Fix x ∈ V x \in V x ∈ V , and recall that M V ( x ) = max y ∈ N V ( x ) ⟨ g , v x y ⟩ M_V(x) = \max_{y \in N_V(x)} \langle g, v_{xy}\rangle M V ( x ) = max y ∈ N V ( x ) ⟨ g , v x y ⟩ . Each v x y v_{xy} v x y is a unit vector, so ⟨ g , v x y ⟩ ∼ N ( 0 , 1 ) \langle g, v_{xy}\rangle\sim N(0,1) ⟨ g , v x y ⟩ ∼ N ( 0 , 1 ) , and x x x has at most n β n^\beta n β neighbors. Hence a union bound and the upper bound of Theorem 12 give P g [ M V ( x ) > a ] ≤ ∣ N V ( x ) ∣ Φ ‾ ( a ) ≤ n β e − a 2 / 2 \mathbb{P}_g[M_V(x) > a] \le|N_V(x)|\overline{\Phi}(a) \le n^\beta e^{-a^2/2} P g [ M V ( x ) > a ] ≤ ∣ N V ( x ) ∣ Φ ( a ) ≤ n β e − a 2 /2 for every a ≥ 0 a \ge0 a ≥ 0 . Choosing a = 2 log ( 2 n β ) a = \sqrt{2\log(2n^\beta)} a = 2 log ( 2 n β ) , we obtain P g [ M V ( x ) ≤ a ] ≥ 1 / 2 \mathbb{P}_g[M_V(x) \le a] \ge1/2 P g [ M V ( x ) ≤ a ] ≥ 1/2 , and hence, by the second claim of Theorem 17 with F = V F = V F = V , p V ( x ) ≥ P g [ M V ( x ) ≤ a ] Φ ‾ ( a / 3 ) ≥ p ∗ p_V(x) \ge\mathbb{P}_g[M_V(x) \le a]\overline{\Phi}(a/\sqrt{3}) \ge p_* p V ( x ) ≥ P g [ M V ( x ) ≤ a ] Φ ( a / 3 ) ≥ p ∗ . By the lower bound of Theorem 12,
p ∗ ≥ c 0 2 ( 1 + 2 3 log ( 2 n β ) ) exp ( − 1 3 log ( 2 n β ) ) = Ω ( n − β / 3 1 + log ( 2 n β ) ) = n − β / 3 − o ( 1 ) . p_* \ge\frac{c_0}{2\left(1+\sqrt{\frac{2}{3}\log(2n^\beta)}\right)}\exp\left(-\frac{1}{3}\log(2n^\beta)\right)
= \Omega\left(\frac{n^{-\beta/3}}{1+\sqrt{\log(2n^\beta)}}\right)
= n^{-\beta/3-o(1)}. p ∗ ≥ 2 ( 1 + 3 2 log ( 2 n β ) ) c 0 exp ( − 3 1 log ( 2 n β ) ) = Ω ( 1 + log ( 2 n β ) n − β /3 ) = n − β /3 − o ( 1 ) .
Hence E g [ ∣ I V ( g ) ∣ ] = ∑ x ∈ V p V ( x ) ≥ n p ∗ \mathbb{E}_g[|I_V(g)|] = \sum_{x\in V}p_V(x) \ge np_* E g [ ∣ I V ( g ) ∣ ] = ∑ x ∈ V p V ( x ) ≥ n p ∗ . Since 0 ≤ ∣ I V ( g ) ∣ ≤ n 0 \le|I_V(g)| \le n 0 ≤ ∣ I V ( g ) ∣ ≤ n , we have n p ∗ ≤ E g [ ∣ I V ( g ) ∣ ] ≤ n p ∗ 2 + n P g [ ∣ I V ( g ) ∣ ≥ n p ∗ / 2 ] np_* \le\mathbb{E}_g[|I_V(g)|] \le\frac{np_*}{2} + n\mathbb{P}_g[|I_V(g)| \ge np_*/2] n p ∗ ≤ E g [ ∣ I V ( g ) ∣ ] ≤ 2 n p ∗ + n P g [ ∣ I V ( g ) ∣ ≥ n p ∗ /2 ] , and therefore P g [ ∣ I V ( g ) ∣ ≥ n p ∗ / 2 ] ≥ p ∗ / 2 \mathbb{P}_g[|I_V(g)| \ge np_*/2] \ge p_*/2 P g [ ∣ I V ( g ) ∣ ≥ n p ∗ /2 ] ≥ p ∗ /2 .
At the first level, Algorithm 2 runs Algorithm 1 R R R times independently on F 0 = V F_0 = V F 0 = V and returns SUCCESS ( I ( i ) ) \operatorname{SUCCESS}(I^{(i)}) SUCCESS ( I ( i ) ) as soon as ∣ I ( i ) ∣ ≥ T |I^{(i)}| \ge T ∣ I ( i ) ∣ ≥ T . Hence the probability that no sample at the first level reaches T T T is at most ( 1 − p ∗ / 2 ) R ≤ exp ( − R p ∗ / 2 ) ≤ 1 / 4 (1-p_*/2)^R \le\exp(-Rp_*/2) \le1/4 ( 1 − p ∗ /2 ) R ≤ exp ( − R p ∗ /2 ) ≤ 1/4 . Thus, with probability at least 3 / 4 3/4 3/4 , the trial returns SUCCESS ( I ) \operatorname{SUCCESS}(I) SUCCESS ( I ) with ∣ I ∣ ≥ T ≥ n p ∗ / 2 − 1 = n 1 − β / 3 − o ( 1 ) |I| \ge T \ge np_*/2 - 1 = n^{1-\beta/3-o(1)} ∣ I ∣ ≥ T ≥ n p ∗ /2 − 1 = n 1 − β /3 − o ( 1 ) . Finally, one trial runs Algorithm 1 at most ℓ R \ell R ℓ R times and R = n β / 3 + o ( 1 ) R = n^{\beta/3+o(1)} R = n β /3 + o ( 1 ) , so it runs in polynomial time. □
Failed Trials and Initial Gaussian Covers
For the rest of this section, we treat the high-degree regime and assume 1 / 3 < β ≤ 1 1/3 < \beta\le1 1/3 < β ≤ 1 . In this subsection, we first prove two lemmas about a failed trial: (1) if Algorithm 2 fails, the deepest core F ℓ F_\ell F ℓ is nonempty, and (2) with probability at least 3 / 4 3/4 3/4 , every surviving vertex x ∈ F j x \in F_j x ∈ F j has small p F j − 1 ( x ) p_{F_{j-1}}(x) p F j − 1 ( x ) , as predicted by p ^ j ( x ) \widehat{p}_j(x) p j ( x ) . We then convert the smallness of p F j − 1 ( x ) p_{F_{j-1}}(x) p F j − 1 ( x ) into a Gaussian cover around a vertex of the deepest core.
Failed Trials
Lemma 19. If every sampled independent set I ( i ) I^{(i)} I ( i ) has size less than T T T at level j j j , then
∣ F j − 1 ∖ F j ∣ < 2 T δ . |F_{j-1} \setminus F_j| < \frac{2T}{\delta}. ∣ F j − 1 ∖ F j ∣ < δ 2 T . In particular, if this holds at all levels j = 1 , … , ℓ j=1,\ldots,\ell j = 1 , … , ℓ , then
∑ j = 1 ℓ ∣ F j − 1 ∖ F j ∣ < 2 ℓ T δ . \sum_{j=1}^{\ell} |F_{j-1} \setminus F_j| < \frac{2\ell T}{\delta}. j = 1 ∑ ℓ ∣ F j − 1 ∖ F j ∣ < δ 2 ℓ T .
Proof. By the definition of p ^ j ( x ) \hat{p}_j(x) p ^ j ( x ) and the assumption that every sampled independent set has size less than T T T , ∑ x ∈ F j − 1 p ^ j ( x ) = R − 1 ∑ i = 1 R ∣ I ( i ) ∣ < T \sum_{x\in F_{j-1}}\hat{p}_j(x)=R^{-1}\sum_{i=1}^{R}|I^{(i)}|<T ∑ x ∈ F j − 1 p ^ j ( x ) = R − 1 ∑ i = 1 R ∣ I ( i ) ∣ < T . Since every deleted vertex satisfies p ^ j ( x ) ≥ δ / 2 \hat{p}_j(x)\geq\delta/2 p ^ j ( x ) ≥ δ /2 , we obtain ∣ F j − 1 ∖ F j ∣ < 2 T δ |F_{j-1}\setminus F_j|<\frac{2T}{\delta} ∣ F j − 1 ∖ F j ∣ < δ 2 T . Summing over j j j gives the second claim. □
Theorem 19 holds for arbitrary δ \delta δ and T T T . To keep the deepest core nonempty, we use their high-degree values in Table 2 .
Lemma 20. If every sampled independent set has size less than T T T at all levels j = 1 , … , ℓ j=1,\ldots,\ell j = 1 , … , ℓ , then F ℓ F_\ell F ℓ is nonempty for sufficiently large n n n .
Proof. Since V = F 0 ⊇ F 1 ⊇ ⋯ ⊇ F ℓ V=F_0\supseteq F_1\supseteq\cdots\supseteq F_\ell V = F 0 ⊇ F 1 ⊇ ⋯ ⊇ F ℓ , we have
∣ F ℓ ∣ = ∣ F 0 ∣ − ∑ j = 1 ℓ ( ∣ F j − 1 ∣ − ∣ F j ∣ ) |F_\ell|=|F_0|-\sum_{j=1}^{\ell}(|F_{j-1}|-|F_j|) ∣ F ℓ ∣ = ∣ F 0 ∣ − j = 1 ∑ ℓ ( ∣ F j − 1 ∣ − ∣ F j ∣ ) = ∣ V ∣ − ∑ j = 1 ℓ ∣ F j − 1 ∖ F j ∣ =|V|-\sum_{j=1}^{\ell}|F_{j-1}\setminus F_j| = ∣ V ∣ − j = 1 ∑ ℓ ∣ F j − 1 ∖ F j ∣ > n − 2 ℓ T δ , >n-\frac{2\ell T}{\delta}, > n − δ 2 ℓ T , where the last inequality is Theorem 19. Recall that δ − 1 = n β / ( 2 + 3 β ) ℓ 500 f ( β ) \delta^{-1}=n^{\beta/(2+3\beta)}\ell^{500f(\beta)} δ − 1 = n β / ( 2 + 3 β ) ℓ 500 f ( β ) , ℓ = ⌈ log n ⌉ \ell=\lceil\log n\rceil ℓ = ⌈ log n ⌉ and T ≤ n δ / ( 100 ℓ 4 ) + 1 T\leq n\delta/(100\ell^{4})+1 T ≤ n δ / ( 100 ℓ 4 ) + 1 . Thus
2 ℓ T δ ≤ n 50 ℓ 3 + 2 ℓ δ = O ( n / ℓ 3 ) + O ( n β / ( 2 + 3 β ) ℓ 500 f ( β ) + 1 ) = o ( n ) . \frac{2\ell T}{\delta}\leq\frac{n}{50\ell^3}+\frac{2\ell}{\delta}=O(n/\ell^3)+O(n^{\beta/(2+3\beta)}\ell^{500f(\beta)+1})=o(n). δ 2 ℓ T ≤ 50 ℓ 3 n + δ 2 ℓ = O ( n / ℓ 3 ) + O ( n β / ( 2 + 3 β ) ℓ 500 f ( β ) + 1 ) = o ( n ) . Hence ∣ F ℓ ∣ > n − o ( n ) ≥ 1 |F_\ell|>n-o(n)\geq1 ∣ F ℓ ∣ > n − o ( n ) ≥ 1 for sufficiently large n n n , completing the proof. □
Next, the sample count R R R of Table 2 is chosen to make the empirical estimates accurate.
Lemma 21. With probability at least 3 4 \frac{3}{4} 4 3 over the samples used to construct the cores, ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ / 8 |\hat{p}_j(x)-p_{F_{j-1}}(x)|\leq\delta/8 ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ /8 for every level j j j at which F j F_j F j is constructed and every vertex x ∈ F j − 1 x\in F_{j-1} x ∈ F j − 1 ; in particular, p F j − 1 ( x ) < δ p_{F_{j-1}}(x)<\delta p F j − 1 ( x ) < δ for every such j j j and every x ∈ F j x\in F_j x ∈ F j .
Proof. By the definition of F j F_j F j , every x ∈ F j x\in F_j x ∈ F j satisfies p ^ j ( x ) < δ / 2 \hat{p}_j(x)<\delta/2 p ^ j ( x ) < δ /2 . Fix any realization of all samples preceding level j j j . This fixes F j − 1 F_{j-1} F j − 1 , and the R R R samples drawn at level j j j are then independent. Hence, for each fixed x ∈ F j − 1 x\in F_{j-1} x ∈ F j − 1 , Hoeffding’s inequality gives
P [ ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ > δ / 8 ] < 2 e − R δ 2 / 32 ≤ 1 4 n ℓ \mathbb{P}[|\hat{p}_j(x)-p_{F_{j-1}}(x)|>\delta/8]<2e^{-R\delta^2/32}\leq\frac{1}{4n\ell} P [ ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ > δ /8 ] < 2 e − R δ 2 /32 ≤ 4 n ℓ 1
by the choice of R R R . Since this bound holds for every realization of the past, a union bound over all levels and vertices gives
P [ ∃ j ∈ { 1 , … , ℓ } , ∃ x ∈ F j − 1 : ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ > δ / 8 ] < 1 4 . \mathbb{P}\left[\exists j \in\{1,\ldots,\ell\},\exists x \in F_{j-1}: \left|\hat{p}_j(x)-p_{F_{j-1}}(x)\right|>\delta/8\right]<\frac{1}{4}. P [ ∃ j ∈ { 1 , … , ℓ } , ∃ x ∈ F j − 1 : p ^ j ( x ) − p F j − 1 ( x ) > δ /8 ] < 4 1 .
Therefore, we have p F j − 1 ( x ) ≤ p ^ j ( x ) + δ / 8 < 5 δ / 8 < δ p_{F_{j-1}}(x)\leq\hat{p}_j(x)+\delta/8<5\delta/8<\delta p F j − 1 ( x ) ≤ p ^ j ( x ) + δ /8 < 5 δ /8 < δ for all levels j j j and all vertices x ∈ F j x\in F_j x ∈ F j . □
The Initial Cover
In the Gaussian cover analysis, we regard the sampled cores as fixed and evaluate probabilities with respect to a fresh Gaussian vector g g g independent of the samples used to construct them.
Definition 22 (Gaussian cover). A finite family of vectors X X X is called an ( h , q ) (h,q) ( h , q ) -cover for a function h : X → R h:X\to\mathbb{R} h : X → R if
P g [ ∃ z ∈ X : ⟨ g , z ⟩ ≥ h ( z ) ] ≥ q . \mathbb{P}_g[\exists z\in X:\langle g,z\rangle\geq h(z)]\geq q. P g [ ∃ z ∈ X : ⟨ g , z ⟩ ≥ h ( z )] ≥ q . In particular, for a constant function h ≡ s h\equiv s h ≡ s , this condition reads
P g [ max z ∈ X ⟨ g , z ⟩ ≥ s ] ≥ q , \mathbb{P}_g\left[\max_{z\in X}\langle g,z\rangle\geq s\right]\geq q, P g [ z ∈ X max ⟨ g , z ⟩ ≥ s ] ≥ q , and we simply call X X X an ( s , q ) (s,q) ( s , q ) -cover with cover threshold s s s and covering probability $q.
Lemma 23. Let F ⊆ V F\subseteq V F ⊆ V and r ∈ F r\in F r ∈ F , and suppose that p F ( r ) ≤ e − t 2 / 2 p_F(r)\leq e^{-t^2/2} p F ( r ) ≤ e − t 2 /2 . Then, for every a ∈ R a\in\mathbb{R} a ∈ R ,
P g [ M F ( r ) > a ] ≥ 1 − e − t 2 / 2 Φ ‾ ( a / 3 ) . \mathbb{P}_g[M_F(r)>a]\geq1-\frac{e^{-t^2/2}}{\overline{\Phi}(a/\sqrt{3})}. P g [ M F ( r ) > a ] ≥ 1 − Φ ( a / 3 ) e − t 2 /2 . Proof. Applying Theorem 17 with x = r x=r x = r gives
P g [ M F ( r ) ≤ a ] ≤ e − t 2 / 2 Φ ‾ ( a / 3 ) . \mathbb{P}_g[M_F(r)\leq a]\leq\frac{e^{-t^2/2}}{\overline{\Phi}(a/\sqrt{3})}. P g [ M F ( r ) ≤ a ] ≤ Φ ( a / 3 ) e − t 2 /2 . Therefore, with probability at least 1 − e − t 2 / 2 / Φ ‾ ( a / 3 ) 1-e^{-t^2/2}/\overline{\Phi}(a/\sqrt{3}) 1 − e − t 2 /2 / Φ ( a / 3 ) , we have M F ( r ) > a M_F(r)>a M F ( r ) > a . □
Fix an arbitrary vertex r ∈ F ℓ r\in F_\ell r ∈ F ℓ of the deepest core, call it the root, and use ρ x \rho_x ρ x and u x u_x u x of Subsection 3.2 for this root. Set W 0 = N F ℓ − 1 ( r ) W_0=N_{F_{\ell-1}}(r) W 0 = N F ℓ − 1 ( r ) , and recall that U 0 = 3 / ( 2 2 ) U_0=3/(2\sqrt{2}) U 0 = 3/ ( 2 2 ) .
Corollary 24. Suppose that p F ℓ − 1 ( r ) ≤ e − t 2 / 2 p_{F_{\ell-1}}(r)\leq e^{-t^2/2} p F ℓ − 1 ( r ) ≤ e − t 2 /2 . Then { u y } y ∈ W 0 \{u_y\}_{y\in W_0} { u y } y ∈ W 0 is a ( U 0 t , 1 − O ( t e − t 2 / 4 ) ) (U_0t,1-O(te^{-t^2/4})) ( U 0 t , 1 − O ( t e − t 2 /4 )) -cover.
Proof. Applying Theorem 23 with F = F ℓ − 1 F=F_{\ell-1} F = F ℓ − 1 , r ∈ F ℓ ⊆ F ℓ − 1 r\in F_\ell\subseteq F_{\ell-1} r ∈ F ℓ ⊆ F ℓ − 1 , and a = 3 / 2 t a=\sqrt{3/2}\,t a = 3/2 t gives
P g [ M F ℓ − 1 ( r ) > 3 / 2 t ] ≥ 1 − e − t 2 / 2 Φ ‾ ( t / 2 ) = 1 − O ( t e − t 2 / 4 ) , \mathbb{P}_g[M_{F_{\ell-1}}(r)>\sqrt{3/2}\,t]\geq1-\frac{e^{-t^2/2}}{\overline{\Phi}(t/\sqrt{2})}=1-O(te^{-t^2/4}), P g [ M F ℓ − 1 ( r ) > 3/2 t ] ≥ 1 − Φ ( t / 2 ) e − t 2 /2 = 1 − O ( t e − t 2 /4 ) , where the last step uses the lower bound of Theorem 12 with u = t / 2 u=t/\sqrt{2} u = t / 2 . Therefore, with probability at least 1 − O ( t e − t 2 / 4 ) 1-O(te^{-t^2/4}) 1 − O ( t e − t 2 /4 ) , we have M F ℓ − 1 ( r ) ≥ 3 / 2 t M_{F_{\ell-1}}(r)\geq\sqrt{3/2}\,t M F ℓ − 1 ( r ) ≥ 3/2 t . Hence { v r v y } y ∈ W 0 \{v_rv_y\}_{y\in W_0} { v r v y } y ∈ W 0 is a ( 3 / 2 t , 1 − O ( t e − t 2 / 4 ) ) (\sqrt{3/2}\,t,1-O(te^{-t^2/4})) ( 3/2 t , 1 − O ( t e − t 2 /4 )) -cover. Moreover, since each vertex y ∈ W 0 y\in W_0 y ∈ W 0 is adjacent to the root r r r , we have u y = 3 2 v r v y u_y=\frac{\sqrt{3}}{2}v_rv_y u y = 2 3 v r v y , so { u y } y ∈ W 0 \{u_y\}_{y\in W_0} { u y } y ∈ W 0 is a ( 3 2 3 / 2 t , 1 − O ( t e − t 2 / 4 ) ) (\frac{\sqrt{3}}{2}\sqrt{3/2}\,t,1-O(te^{-t^2/4})) ( 2 3 3/2 t , 1 − O ( t e − t 2 /4 )) -cover, and 3 2 3 / 2 = U 0 \frac{\sqrt{3}}{2}\sqrt{3/2}=U_0 2 3 3/2 = U 0 . This completes the proof. □
Lifting to the Preceding Level
The lifting argument involves two cardinality bounds: the total number of vertices n n n and the maximum degree n β n^\beta n β . In terms of the quantities H 0 H_0 H 0 and D D D of Table 2 , these bounds read n = e H 0 t 2 / 2 n=e^{H_0t^2/2} n = e H 0 t 2 /2 and ∣ N F ( x ) ∣ ≤ n β = e D t 2 / 2 |N_F(x)|\le n^\beta=e^{Dt^2/2} ∣ N F ( x ) ∣ ≤ n β = e D t 2 /2 for every F ⊆ V F\subseteq V F ⊆ V and x ∈ V x\in V x ∈ V . In this subsection, we prove the single-step lifting lemma (Theorem 25). Given a ( U t , q ) (Ut,q) ( U t , q ) -cover { u x } x ∈ W \{u_x\}_{x\in W} { u x } x ∈ W with W W W a subset of the level-j j j core F j F_j F j , we prove that { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( U + Δ U ) t , q − Δ q ) ((U+\Delta U)t,q-\Delta q) (( U + Δ U ) t , q − Δ q ) -cover for some W ′ ⊆ F j − 1 W'\subseteq F_{j-1} W ′ ⊆ F j − 1 and appropriate Δ U \Delta U Δ U and Δ q \Delta q Δ q . More precisely, we can take Δ U t = Θ ( w t ) = Θ ( log t t ) = Θ ( log log n log n ) \Delta Ut=\Theta(wt)=\Theta\left(\frac{\log t}{t}\right)=\Theta\left(\frac{\log\log n}{\sqrt{\log n}}\right) Δ U t = Θ ( w t ) = Θ ( t l o g t ) = Θ ( l o g n l o g l o g n ) and Δ q = O ( t − 18 ) = O ( ( log n ) − 9 ) \Delta q=O(t^{-18})=O((\log n)^{-9}) Δ q = O ( t − 18 ) = O (( log n ) − 9 ) .
The positive drift in U U U is sufficiently large for U U U to eventually exceed H \sqrt{H} H , while the accumulated loss in the covering probability remains negligible, leading to a contradiction with the Gaussian tail bound.
Lemma 25 (Single-step lifting). Let 1 ≤ j ≤ ℓ 1\le j\le\ell 1 ≤ j ≤ ℓ , and let W ⊆ F j W\subseteq F_j W ⊆ F j be fixed independently of the Gaussian g g g , and suppose that { u x } x ∈ W \{u_x\}_{x\in W} { u x } x ∈ W is a ( U t , q ) (Ut,q) ( U t , q ) -cover for some 0 < U ≤ H 0<U\le\sqrt{H} 0 < U ≤ H . Assume that p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 for every x ∈ W x\in W x ∈ W , and set W ′ = N F j − 1 ( W ) W'=N_{F_{j-1}}(W) W ′ = N F j − 1 ( W ) . Then, for sufficiently large t t t , { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( U + 40 w ) t , q − t − 18 ) ((U+40w)t,q-t^{-18}) (( U + 40 w ) t , q − t − 18 ) -cover.
For x ∈ F j x\in F_j x ∈ F j , we begin with an observation of the relation between the parent Gaussian projection ⟨ g , u x ⟩ \langle g,u_x\rangle ⟨ g , u x ⟩ and the child Gaussian projection ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ for y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) . We use the decomposition in Equation (4) of Subsection 3.2. We write g ⊥ g_\perp g ⊥ for the orthogonal projection of g g g onto span ( v r , v x ) ⊥ \operatorname{span}(v_r,v_x)^\perp span ( v r , v x ) ⊥ , so that ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_\perp,w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ .
By the first equality of Equation (6) , we have
⟨ g , u y ⟩ = − ( 1 2 + 3 2 ξ y ρ x ∥ u x ∥ ) ⟨ g , u x ⟩ + 3 2 ⟨ g , w y ⟩ . \langle g,u_y\rangle=-\left(\frac{1}{2}+\frac{\sqrt{3}}{2}\frac{\xi_y\rho_x}{\lVert u_x\rVert}\right)\langle g,u_x\rangle+\frac{\sqrt{3}}{2}\langle g,w_y\rangle. ⟨ g , u y ⟩ = − ( 2 1 + 2 3 ∥ u x ∥ ξ y ρ x ) ⟨ g , u x ⟩ + 2 3 ⟨ g , w y ⟩ .
Since t 2 = 2 β 2 + 3 β log n + 1000 f ( β ) log ℓ t^2=\frac{2\beta}{2+3\beta}\log n+1000f(\beta)\log\ell t 2 = 2 + 3 β 2 β log n + 1000 f ( β ) log ℓ , the parameters D D D , H 0 H_0 H 0 and H H H satisfy D → 2 + 3 β D\to2+3\beta D → 2 + 3 β and H 0 , H → ( 2 + 3 β ) / β H_0,H\to(2+3\beta)/\beta H 0 , H → ( 2 + 3 β ) / β as n → ∞ n\to\infty n → ∞ . Since β > 1 / 3 \beta>1/3 β > 1/3 , this gives D > 3 D>3 D > 3 and H > 1 H>1 H > 1 for sufficiently large n n n . Define ν = 1 2 ( D − 3 ) / D \nu=\frac{1}{2}\sqrt{(D-3)/D} ν = 2 1 ( D − 3 ) / D , and let
ψ ( s ) = { 3 1 − s 2 , s ≤ ν , 3 ( 1 − ν s ) 1 − ν 2 , s ≥ ν . \psi(s)=
\begin{cases}
\sqrt{3}\sqrt{1-s^2}, & s\le\nu,\\
\frac{\sqrt{3}(1-\nu s)}{\sqrt{1-\nu^2}}, & s\ge\nu.
\end{cases} ψ ( s ) = { 3 1 − s 2 , 1 − ν 2 3 ( 1 − ν s ) , s ≤ ν , s ≥ ν .
Thus, ψ \psi ψ agrees with the curve s ↦ 3 1 − s 2 s\mapsto\sqrt{3}\sqrt{1-s^2} s ↦ 3 1 − s 2 on [ 0 , ν ] [0,\nu] [ 0 , ν ] and with its tangent line at s = ν s=\nu s = ν on [ ν , ∞ ) [\nu,\infty) [ ν , ∞ ) .
Here, assuming Theorem 26 and Theorem 27 below, we prove Theorem 25. The first lemma provides the probabilistic guarantee for transmission to a neighboring vertex, while the second ensures a uniform positive increase in the cover threshold.
Lemma 26. Let 1 ≤ j ≤ ℓ 1\le j\le\ell 1 ≤ j ≤ ℓ , and let x ∈ F j x\in F_j x ∈ F j satisfy ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 . Let 0 < L < H 0<L<\sqrt{H} 0 < L < H and ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] . Then, for sufficiently large t t t , over the Gaussian coordinates orthogonal to span ( v r , v x ) \operatorname{span}(v_r,v_x) span ( v r , v x ) ,
P [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − L 2 ( 1 − ρ 2 ) ) − 40 ( 1 + 20 f ( β ) ) w ) t ] ≥ Φ ( H − L 2 t ) . \mathbb{P}\left[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle+L\rho\xi_y t\ge\left(\psi\left(\sqrt{H-L^2(1-\rho^2)}\right)-40(1+20f(\beta))w\right)t\right]
\ge\Phi\left(\sqrt{H-L^2}\,t\right). P [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − L 2 ( 1 − ρ 2 ) ) − 40 ( 1 + 20 f ( β )) w ) t ] ≥ Φ ( H − L 2 t ) .
Lemma 27. For sufficiently large n n n ,
3 2 ψ ( H − U 2 ) ≥ U 2 + 800 f ( β ) w \frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^2}\right)\ge\frac{U}{2}+800f(\beta)w 2 3 ψ ( H − U 2 ) ≥ 2 U + 800 f ( β ) w for every 0 ≤ U ≤ H 0\le U\le\sqrt{H} 0 ≤ U ≤ H .
Proof of Theorem 25. By Gaussian symmetry, the ( U t , q ) (Ut,q) ( U t , q ) -cover assumption also gives
P g [ ∃ x ∈ W : ⟨ g , u x ⟩ ≤ − U t ] ≥ q . \mathbb{P}_{g}\left[\exists x\in W:\langle g,u_x\rangle\le-Ut\right]\ge q. P g [ ∃ x ∈ W : ⟨ g , u x ⟩ ≤ − U t ] ≥ q .
Fix x ∈ W x\in W x ∈ W and let E x \mathcal{E}_x E x denote the event ⟨ g , u x ⟩ ≤ − U t \langle g,u_x\rangle\le-Ut ⟨ g , u x ⟩ ≤ − U t . If ∥ u x ∥ = 0 \lVert u_x\rVert=0 ∥ u x ∥ = 0 , then E x \mathcal{E}_x E x cannot occur, so assume ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 . Set Z x = − ⟨ g , u x / ∥ u x ∥ ⟩ Z_x=-\langle g,u_x/\lVert u_x\rVert\rangle Z x = − ⟨ g , u x / ∥ u x ∥⟩ , which is a standard Gaussian, and let
E x b a d = E x ∩ { ∀ y ∈ N F j − 1 ( x ) : ⟨ g , u y ⟩ < ( U + 40 w ) t } . \mathcal{E}_x^{\mathrm{bad}}=\mathcal{E}_x\cap\left\{\forall y\in N_{F_{j-1}}(x):\langle g,u_y\rangle<(U+40w)t\right\}. E x bad = E x ∩ { ∀ y ∈ N F j − 1 ( x ) : ⟨ g , u y ⟩ < ( U + 40 w ) t } .
On E x \mathcal{E}_x E x , we have Z x ≥ U t / ∥ u x ∥ Z_x\ge Ut/\lVert u_x\rVert Z x ≥ U t / ∥ u x ∥ . Hence P g [ E x b a d ∣ Z x = z ] = 0 \mathbb{P}_g[\mathcal{E}_x^{\mathrm{bad}}\mid Z_x=z]=0 P g [ E x bad ∣ Z x = z ] = 0 for z < U t / ∥ u x ∥ z<Ut/\lVert u_x\rVert z < U t / ∥ u x ∥ .
Now fix z z z with U t / ∥ u x ∥ ≤ z < H t Ut/\lVert u_x\rVert\le z<\sqrt{H}t U t / ∥ u x ∥ ≤ z < H t and condition on Z x = z Z_x=z Z x = z . Since u x / ∥ u x ∥ ∈ span ( v r , v x ) u_x/\lVert u_x\rVert\in\operatorname{span}(v_r,v_x) u x / ∥ u x ∥ ∈ span ( v r , v x ) , the Gaussian coordinates orthogonal to span ( v r , v x ) \operatorname{span}(v_r,v_x) span ( v r , v x ) are unaffected by this conditioning.
Since ∥ u x ∥ ≤ 1 \lVert u_x\rVert\le1 ∥ u x ∥ ≤ 1 , we have U ≤ ∥ u x ∥ z / t < H U\le\lVert u_x\rVert z/t<\sqrt{H} U ≤ ∥ u x ∥ z / t < H .
Applying Theorem 26 with L = z / t L=z/t L = z / t and ρ = ρ x \rho=\rho_x ρ = ρ x , where L 1 − ρ 2 = ∥ u x ∥ z / t L\sqrt{1-\rho^2}=\lVert u_x\rVert z/t L 1 − ρ 2 = ∥ u x ∥ z / t , gives
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + ρ ξ y z ≥ ( ψ ( H − ∥ u x ∥ 2 z 2 t 2 ) − 40 ( 1 + 20 f ( β ) ) w ) t | Z x = z ] ≥ Φ ( H t 2 − z 2 ) . \mathbb{P}_g\left[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle+\rho\xi_yz\ge\left(\psi\left(\sqrt{H-\frac{\lVert u_x\rVert^2z^2}{t^2}}\right)-40(1+20f(\beta))w\right)t\ \middle|\ Z_x=z\right]
\ge\Phi\left(\sqrt{Ht^2-z^2}\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + ρ ξ y z ≥ ( ψ ( H − t 2 ∥ u x ∥ 2 z 2 ) − 40 ( 1 + 20 f ( β )) w ) t Z x = z ] ≥ Φ ( H t 2 − z 2 ) .
When the event above happens,
⟨ g , u y ⟩ = ∥ u x ∥ z 2 + 3 2 ( ρ ξ y z + ⟨ g , w y ⟩ ) ≥ U t 2 + 3 2 ψ ( H − ∥ u x ∥ 2 z 2 t 2 ) t − 20 3 ( 1 + 20 f ( β ) ) w t ≥ ( U + 800 f ( β ) w − 20 3 ( 1 + 20 f ( β ) ) w ) t ≥ ( U + 40 w ) t , \begin{aligned}
\langle g,u_y\rangle
&=\frac{\lVert u_x\rVert z}{2}+\frac{\sqrt{3}}{2}\left(\rho\xi_yz+\langle g,w_y\rangle\right)\\
&\ge\frac{Ut}{2}+\frac{\sqrt{3}}{2}\psi\left(\sqrt{H-\frac{\lVert u_x\rVert^2z^2}{t^2}}\right)t-20\sqrt{3}(1+20f(\beta))wt\\
&\ge\left(U+800f(\beta)w-20\sqrt{3}(1+20f(\beta))w\right)t\\
&\ge(U+40w)t,
\end{aligned} ⟨ g , u y ⟩ = 2 ∥ u x ∥ z + 2 3 ( ρ ξ y z + ⟨ g , w y ⟩ ) ≥ 2 U t + 2 3 ψ ( H − t 2 ∥ u x ∥ 2 z 2 ) t − 20 3 ( 1 + 20 f ( β )) w t ≥ ( U + 800 f ( β ) w − 20 3 ( 1 + 20 f ( β )) w ) t ≥ ( U + 40 w ) t ,
where the second inequality is Theorem 27 applied at ∥ u x ∥ z / t \lVert u_x\rVert z/t ∥ u x ∥ z / t , and the last one follows from 800 f ( β ) − 20 3 ( 1 + 20 f ( β ) ) = ( 800 − 400 3 ) f ( β ) − 20 3 ≥ 40 800f(\beta)-20\sqrt{3}(1+20f(\beta))=(800-400\sqrt{3})f(\beta)-20\sqrt{3}\ge40 800 f ( β ) − 20 3 ( 1 + 20 f ( β )) = ( 800 − 400 3 ) f ( β ) − 20 3 ≥ 40 , which holds since f ( β ) ≥ 1 / 2 f(\beta)\ge1/\sqrt{2} f ( β ) ≥ 1/ 2 .
Therefore
P g [ E x b a d ∣ Z x = z ] ≤ P g [ ∀ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + ρ ξ y z < ( ψ ( H − ∥ u x ∥ 2 z 2 t 2 ) − 40 ( 1 + 20 f ( β ) ) w ) t | Z x = z ] ≤ Φ ‾ ( H t 2 − z 2 ) . \begin{aligned}
\mathbb{P}_g[\mathcal{E}_x^{\mathrm{bad}}\mid Z_x=z] \\
&\le\mathbb{P}_g\left[\forall y\in N_{F_{j-1}}(x):\langle g,w_y\rangle+\rho\xi_yz<\left(\psi\left(\sqrt{H-\frac{\lVert u_x\rVert^2z^2}{t^2}}\right)-40(1+20f(\beta))w\right)t\ \middle|\ Z_x=z\right] \\
&\le\overline{\Phi}\left(\sqrt{Ht^2-z^2}\right).
\end{aligned} P g [ E x bad ∣ Z x = z ] ≤ P g [ ∀ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + ρ ξ y z < ( ψ ( H − t 2 ∥ u x ∥ 2 z 2 ) − 40 ( 1 + 20 f ( β )) w ) t Z x = z ] ≤ Φ ( H t 2 − z 2 ) .
Integrating over z z z and bounding P g [ Z x ≥ H t ] \mathbb{P}_{g}[Z_x \ge\sqrt{H}t] P g [ Z x ≥ H t ] by Φ ‾ ( H t ) \overline{\Phi}(\sqrt{H}t) Φ ( H t ) ,
P g [ E x b a d ] ≤ Φ ‾ ( H t ) + ∫ 0 H t ϕ ( z ) Φ ‾ ( H t 2 − z 2 ) d z ≤ e − H t 2 / 2 + ∫ 0 H t 1 2 π e − z 2 / 2 e − ( H t 2 − z 2 ) / 2 d z = ( 1 + H t 2 π ) e − H t 2 / 2 , \begin{aligned}
\mathbb{P}_{g}[\mathcal{E}_{x}^{\mathrm{bad}}] \le\overline{\Phi}(\sqrt{H}t) + \int_{0}^{\sqrt{H}t} \phi(z)\overline{\Phi}\left(\sqrt{Ht^{2}-z^{2}}\right)\,dz \\
&\le e^{-Ht^{2}/2}+\int_{0}^{\sqrt{H}t}\frac{1}{\sqrt{2\pi}}e^{-z^{2}/2}e^{-(Ht^{2}-z^{2})/2}\,dz \\
&=\left(1+\frac{\sqrt{H}t}{\sqrt{2\pi}}\right)e^{-Ht^{2}/2},
\end{aligned} P g [ E x bad ] ≤ Φ ( H t ) + ∫ 0 H t ϕ ( z ) Φ ( H t 2 − z 2 ) d z ≤ e − H t 2 /2 + ∫ 0 H t 2 π 1 e − z 2 /2 e − ( H t 2 − z 2 ) /2 d z = ( 1 + 2 π H t ) e − H t 2 /2 ,
where we used the Gaussian tail bound Φ ‾ ( u ) ≤ e − u 2 / 2 \overline{\Phi}(u)\le e^{-u^{2}/2} Φ ( u ) ≤ e − u 2 /2 of Theorem 12.
Taking a union bound over x ∈ W x\in W x ∈ W and using ∣ W ∣ ≤ n = e H 0 t 2 / 2 |W|\le n=e^{H_{0}t^{2}/2} ∣ W ∣ ≤ n = e H 0 t 2 /2 ,
P g [ ∃ x ∈ W : E x b a d ] ≤ ( 1 + H t 2 π ) e − ( H − H 0 ) t 2 / 2 = ( 1 + H t 2 π ) t − 20 ≤ t − 18 \mathbb{P}_{g}\left[\exists x\in W:\mathcal{E}_{x}^{\mathrm{bad}}\right]\le\left(1+\frac{\sqrt{H}t}{\sqrt{2\pi}}\right)e^{-(H-H_{0})t^{2}/2}=\left(1+\frac{\sqrt{H}t}{\sqrt{2\pi}}\right)t^{-20}\le t^{-18} P g [ ∃ x ∈ W : E x bad ] ≤ ( 1 + 2 π H t ) e − ( H − H 0 ) t 2 /2 = ( 1 + 2 π H t ) t − 20 ≤ t − 18
for sufficiently large t t t , where we used H − H 0 = 40 w H-H_{0}=40w H − H 0 = 40 w , w t 2 = log t wt^{2}=\log t w t 2 = log t , and H = O ( 1 ) H=O(1) H = O ( 1 ) .
Let
E = ⋃ x ∈ W E x , E ′ = { max y ∈ W ′ ⟨ g , u y ⟩ ≥ ( U + 40 w ) t } . \mathcal{E}=\bigcup_{x\in W}\mathcal{E}_{x},\qquad\mathcal{E}'=\left\{\max_{y\in W'}\langle g,u_y\rangle\ge(U+40w)t\right\}. E = x ∈ W ⋃ E x , E ′ = { y ∈ W ′ max ⟨ g , u y ⟩ ≥ ( U + 40 w ) t } .
Then
E ∖ E ′ ⊆ ⋃ x ∈ W E x b a d . \mathcal{E}\setminus\mathcal{E}'\subseteq\bigcup_{x\in W}\mathcal{E}_{x}^{\mathrm{bad}}. E ∖ E ′ ⊆ x ∈ W ⋃ E x bad .
Therefore
P g [ E ′ ] ≥ P g [ E ] − P g [ ⋃ x ∈ W E x b a d ] ≥ q − t − 18 , \mathbb{P}_{g}[\mathcal{E}']\ge\mathbb{P}_{g}[\mathcal{E}]-\mathbb{P}_{g}\left[\bigcup_{x\in W}\mathcal{E}_{x}^{\mathrm{bad}}\right]\ge q-t^{-18}, P g [ E ′ ] ≥ P g [ E ] − P g [ x ∈ W ⋃ E x bad ] ≥ q − t − 18 ,
which shows that { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( U + 40 w ) t , q − t − 18 ) ((U+40w)t,q-t^{-18}) (( U + 40 w ) t , q − t − 18 ) -cover.
First, we prove Theorem 27.
Proof of Theorem 27. Set
ζ = 3 D D + 1 − H ( D − 2 ) D + 1 . \zeta=\sqrt{\frac{3D}{D+1}}-\sqrt{\frac{H(D-2)}{D+1}}. ζ = D + 1 3 D − D + 1 H ( D − 2 ) .
We first show that
3 2 ψ ( H − U 2 ) − U 2 ≥ ζ \frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^{2}}\right)-\frac{U}{2}\ge\zeta 2 3 ψ ( H − U 2 ) − 2 U ≥ ζ
for every 0 ≤ U ≤ H 0\le U\le\sqrt{H} 0 ≤ U ≤ H . Put s = H − U 2 s=\sqrt{H-U^{2}} s = H − U 2 , so that U 2 + s 2 = H U^{2}+s^{2}=H U 2 + s 2 = H .
First suppose that s ≤ ν s\le\nu s ≤ ν . Then
3 2 ψ ( s ) − U 2 = 3 2 1 − s 2 − 1 2 H − s 2 . \frac{\sqrt{3}}{2}\psi(s)-\frac{U}{2}=\frac{3}{2}\sqrt{1-s^{2}}-\frac{1}{2}\sqrt{H-s^{2}}. 2 3 ψ ( s ) − 2 U = 2 3 1 − s 2 − 2 1 H − s 2 .
The right-hand side is decreasing in s ∈ [ 0 , ν ] s\in[0,\nu] s ∈ [ 0 , ν ] , since H > 1 H>1 H > 1 gives
− 3 s 2 1 − s 2 + s 2 H − s 2 ≤ 0. -\frac{3s}{2\sqrt{1-s^{2}}}+\frac{s}{2\sqrt{H-s^{2}}}\le0. − 2 1 − s 2 3 s + 2 H − s 2 s ≤ 0.
Hence
3 2 ψ ( s ) − U 2 ≥ 3 2 ψ ( ν ) − H − ν 2 2 . \frac{\sqrt{3}}{2}\psi(s)-\frac{U}{2}\geq\frac{\sqrt{3}}{2}\psi(\nu)-\frac{\sqrt{H-\nu^{2}}}{2}. 2 3 ψ ( s ) − 2 U ≥ 2 3 ψ ( ν ) − 2 H − ν 2 .
The right-hand side is the value of 3 2 ψ ( s ) − U 2 \frac{\sqrt{3}}{2}\psi(s)-\frac{U}{2} 2 3 ψ ( s ) − 2 U at s = ν s=\nu s = ν . Since the two branches of ψ \psi ψ agree at s = ν s=\nu s = ν , this point is also covered by the second case below. Therefore, the first case follows from the second.
Next, suppose that s ≥ ν s\geq\nu s ≥ ν . Then
3 2 ψ ( s ) − U 2 = 3 2 1 − ν 2 − ( U 2 + 3 ν 2 1 − ν 2 s ) . \frac{\sqrt{3}}{2}\psi(s)-\frac{U}{2}=\frac{3}{2\sqrt{1-\nu^{2}}}-\left(\frac{U}{2}+\frac{3\nu}{2\sqrt{1-\nu^{2}}}s\right). 2 3 ψ ( s ) − 2 U = 2 1 − ν 2 3 − ( 2 U + 2 1 − ν 2 3 ν s ) .
so by Cauchy–Schwarz and U 2 + s 2 = H U^{2}+s^{2}=H U 2 + s 2 = H ,
3 2 ψ ( s ) − U 2 ≥ 3 2 1 − ν 2 − H ( 1 4 + 9 ν 2 4 ( 1 − ν 2 ) ) = ζ , \frac{\sqrt{3}}{2}\psi(s)-\frac{U}{2}\geq\frac{3}{2\sqrt{1-\nu^{2}}}-\sqrt{H\left(\frac{1}{4}+\frac{9\nu^{2}}{4(1-\nu^{2})}\right)}=\zeta, 2 3 ψ ( s ) − 2 U ≥ 2 1 − ν 2 3 − H ( 4 1 + 4 ( 1 − ν 2 ) 9 ν 2 ) = ζ ,
where we used ν 2 = D − 3 4 D \nu^{2}=\frac{D-3}{4D} ν 2 = 4 D D − 3 , hence 1 − ν 2 = 3 ( D + 1 ) 4 D 1-\nu^{2}=\frac{3(D+1)}{4D} 1 − ν 2 = 4 D 3 ( D + 1 ) , to get
3 2 1 − ν 2 = 3 D D + 1 , 1 4 + 9 ν 2 4 ( 1 − ν 2 ) = D − 2 D + 1 . \frac{3}{2\sqrt{1-\nu^{2}}}=\sqrt{\frac{3D}{D+1}},\qquad\frac{1}{4}+\frac{9\nu^{2}}{4(1-\nu^{2})}=\frac{D-2}{D+1}. 2 1 − ν 2 3 = D + 1 3 D , 4 1 + 4 ( 1 − ν 2 ) 9 ν 2 = D + 1 D − 2 .
Thus, 3 2 ψ ( H − U 2 ) − U 2 ≥ ζ \frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^{2}}\right)-\frac{U}{2}\geq\zeta 2 3 ψ ( H − U 2 ) − 2 U ≥ ζ holds.
It remains to lower-bound ζ \zeta ζ . By the definitions of D D D , H 0 H_{0} H 0 , t t t and t 2 = 2 β 2 + 3 β log n + 1000 f ( β ) log ℓ t^{2}=\frac{2\beta}{2+3\beta}\log n+1000f(\beta)\log\ell t 2 = 2 + 3 β 2 β log n + 1000 f ( β ) log ℓ ,
3 D − H 0 ( D − 2 ) = 2000 f ( β ) ( 2 + 3 β ) log n log ℓ t 4 . 3D-H_{0}(D-2)=\frac{2000f(\beta)(2+3\beta)\log n\log\ell}{t^{4}}. 3 D − H 0 ( D − 2 ) = t 4 2000 f ( β ) ( 2 + 3 β ) log n log ℓ .
Hence
3 D − H 0 ( D − 2 ) ( D + 1 ) w = 2000 f ( β ) ( 2 + 3 β ) log n log ℓ ( 2 β log n + t 2 ) log t ⟶ 2000 f ( β ) ( 2 + 3 β ) 2 3 β ( 1 + β ) , \frac{3D-H_{0}(D-2)}{(D+1)w}=\frac{2000f(\beta)(2+3\beta)\log n\log\ell}{(2\beta\log n+t^{2})\log t}\longrightarrow\frac{2000f(\beta)(2+3\beta)^{2}}{3\beta(1+\beta)}, ( D + 1 ) w 3 D − H 0 ( D − 2 ) = ( 2 β log n + t 2 ) log t 2000 f ( β ) ( 2 + 3 β ) log n log ℓ ⟶ 3 β ( 1 + β ) 2000 f ( β ) ( 2 + 3 β ) 2 ,
where we used t 2 log n → 2 β 2 + 3 β \frac{t^{2}}{\log n}\to\frac{2\beta}{2+3\beta} l o g n t 2 → 2 + 3 β 2 β and log ℓ log t → 2 \frac{\log\ell}{\log t}\to2 l o g t l o g ℓ → 2 .
Since H = H 0 + 40 w H=H_{0}+40w H = H 0 + 40 w and D − 2 D + 1 → β 1 + β \frac{D-2}{D+1}\to\frac{\beta}{1+\beta} D + 1 D − 2 → 1 + β β ,
3 D − H ( D − 2 ) ( D + 1 ) w ⟶ 2000 f ( β ) ( 2 + 3 β ) 2 3 β ( 1 + β ) − 40 β 1 + β . \frac{3D-H(D-2)}{(D+1)w}\longrightarrow\frac{2000f(\beta)(2+3\beta)^{2}}{3\beta(1+\beta)}-\frac{40\beta}{1+\beta}. ( D + 1 ) w 3 D − H ( D − 2 ) ⟶ 3 β ( 1 + β ) 2000 f ( β ) ( 2 + 3 β ) 2 − 1 + β 40 β .
Therefore, rationalizing ζ \zeta ζ ,
ζ w = 3 D − H ( D − 2 ) ( D + 1 ) w 3 D D + 1 + H ( D − 2 ) D + 1 ⟶ 1 2 1 + β 2 + 3 β ( 2000 f ( β ) ( 2 + 3 β ) 2 3 β ( 1 + β ) − 40 β 1 + β ) > 800 f ( β ) , \frac{\zeta}{w}=\frac{\frac{3D-H(D-2)}{(D+1)w}}{\sqrt{\frac{3D}{D+1}}+\sqrt{\frac{H(D-2)}{D+1}}}\longrightarrow\frac{1}{2}\sqrt{\frac{1+\beta}{2+3\beta}}\left(\frac{2000f(\beta)(2+3\beta)^{2}}{3\beta(1+\beta)}-\frac{40\beta}{1+\beta}\right)>800f(\beta), w ζ = D + 1 3 D + D + 1 H ( D − 2 ) ( D + 1 ) w 3 D − H ( D − 2 ) ⟶ 2 1 2 + 3 β 1 + β ( 3 β ( 1 + β ) 2000 f ( β ) ( 2 + 3 β ) 2 − 1 + β 40 β ) > 800 f ( β ) ,
where we used 3 D D + 1 → 2 + 3 β 1 + β \frac{3D}{D+1}\to\frac{2+3\beta}{1+\beta} D + 1 3 D → 1 + β 2 + 3 β and H ( D − 2 ) D + 1 → 2 + 3 β 1 + β \frac{H(D-2)}{D+1}\to\frac{2+3\beta}{1+\beta} D + 1 H ( D − 2 ) → 1 + β 2 + 3 β . For the last inequality, note that for β ∈ ( 1 / 3 , 1 ] \beta\in(1/3,1] β ∈ ( 1/3 , 1 ] we have 2 + 3 β β ≥ 5 \frac{2+3\beta}{\beta}\geq5 β 2 + 3 β ≥ 5 , 2 + 3 β ≥ 2 2+3\beta\geq2 2 + 3 β ≥ 2 and 1 + β ≤ 2 1+\beta\leq2 1 + β ≤ 2 , so
1 2 1 + β 2 + 3 β ⋅ 2000 f ( β ) ( 2 + 3 β ) 2 3 β ( 1 + β ) = 1000 f ( β ) 3 ⋅ 2 + 3 β β ⋅ 2 + 3 β 1 + β ≥ 5000 f ( β ) 3 , \frac{1}{2}\sqrt{\frac{1+\beta}{2+3\beta}}\cdot\frac{2000f(\beta)(2+3\beta)^{2}}{3\beta(1+\beta)}=\frac{1000f(\beta)}{3}\cdot\frac{2+3\beta}{\beta}\cdot\sqrt{\frac{2+3\beta}{1+\beta}}\geq\frac{5000f(\beta)}{3}, 2 1 2 + 3 β 1 + β ⋅ 3 β ( 1 + β ) 2000 f ( β ) ( 2 + 3 β ) 2 = 3 1000 f ( β ) ⋅ β 2 + 3 β ⋅ 1 + β 2 + 3 β ≥ 3 5000 f ( β ) ,
while 1 2 1 + β 2 + 3 β ⋅ 40 β 1 + β ≤ 20 \frac{1}{2}\sqrt{\frac{1+\beta}{2+3\beta}}\cdot\frac{40\beta}{1+\beta}\le20 2 1 2 + 3 β 1 + β ⋅ 1 + β 40 β ≤ 20 . Since f ( β ) ≥ 1 / 2 f(\beta)\ge1/\sqrt{2} f ( β ) ≥ 1/ 2 , the limit is therefore at least
5000 f ( β ) 3 − 20 = 800 f ( β ) + 2600 f ( β ) 3 − 20 > 800 f ( β ) . \frac{5000f(\beta)}{3}-20=800f(\beta)+\frac{2600f(\beta)}{3}-20>800f(\beta). 3 5000 f ( β ) − 20 = 800 f ( β ) + 3 2600 f ( β ) − 20 > 800 f ( β ) .
Thus ζ ≥ 800 f ( β ) w \zeta\ge800f(\beta)w ζ ≥ 800 f ( β ) w for sufficiently large n n n . Hence
3 2 ψ ( H − U 2 ) ≥ U 2 + 800 f ( β ) w , \frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^2}\right)\ge\frac{U}{2}+800f(\beta)w, 2 3 ψ ( H − U 2 ) ≥ 2 U + 800 f ( β ) w ,
which proves the claim.
The remaining part of this subsection is devoted to proving Theorem 26.
Ehrhard Extrapolation
To prove Theorem 26, we use Ehrhard’s inequality (Theorem 28). Given a convex set K 0 K_0 K 0 of small Gaussian measure and a convex set K 1 K_1 K 1 of large Gaussian measure, it bounds the Gaussian measure of the extrapolated set K − θ K_{-\theta} K − θ , which is even smaller (Theorem 29). This allows us to show that the transverse Gaussian projection of at least one neighbor of the parent is sufficiently large.
Theorem 28 (Ehrhard’s inequality [19 ] ). If K , K ′ K,K' K , K ′ are convex subsets of a Euclidean space and 0 ≤ λ ≤ 1 0\le\lambda\le1 0 ≤ λ ≤ 1 , then
Φ − 1 γ ( ( 1 − λ ) K + λ K ′ ) ≥ ( 1 − λ ) Φ − 1 γ ( K ) + λ Φ − 1 γ ( K ′ ) . \Phi^{-1}\gamma\left((1-\lambda)K+\lambda K'\right)\ge(1-\lambda)\Phi^{-1}\gamma(K)+\lambda\Phi^{-1}\gamma(K'). Φ − 1 γ ( ( 1 − λ ) K + λ K ′ ) ≥ ( 1 − λ ) Φ − 1 γ ( K ) + λ Φ − 1 γ ( K ′ ) . Here ( 1 − λ ) K + λ K ′ = { ( 1 − λ ) z + λ z ′ : z ∈ K , z ′ ∈ K ′ } (1-\lambda)K+\lambda K'=\{(1-\lambda)z+\lambda z':z\in K,\ z'\in K'\} ( 1 − λ ) K + λ K ′ = {( 1 − λ ) z + λ z ′ : z ∈ K , z ′ ∈ K ′ } . Endpoint and zero-measure cases are interpreted by limits.
Lemma 29. For a finite set Y Y Y and its corresponding vectors { z y } y ∈ Y \{z_y\}_{y\in Y} { z y } y ∈ Y and real thresholds { A y } y ∈ Y , { B y } y ∈ Y \{A_y\}_{y\in Y},\{B_y\}_{y\in Y} { A y } y ∈ Y , { B y } y ∈ Y , let
K 0 = { g : ⟨ g , z y ⟩ ≤ A y ∀ y ∈ Y } , K_0=\{g:\langle g,z_y\rangle\le A_y\ \forall y\in Y\}, K 0 = { g : ⟨ g , z y ⟩ ≤ A y ∀ y ∈ Y } , K 1 = { g : ⟨ g , z y ⟩ ≤ B y ∀ y ∈ Y } , K_1=\{g:\langle g,z_y\rangle\le B_y\ \forall y\in Y\}, K 1 = { g : ⟨ g , z y ⟩ ≤ B y ∀ y ∈ Y } , K − θ = { g : ⟨ g , z y ⟩ ≤ ( 1 + θ ) A y − θ B y ∀ y ∈ Y } . K_{-\theta}=\{g:\langle g,z_y\rangle\le(1+\theta)A_y-\theta B_y\ \forall y\in Y\}. K − θ = { g : ⟨ g , z y ⟩ ≤ ( 1 + θ ) A y − θ B y ∀ y ∈ Y } . If γ ( K 0 ) ≤ Φ ‾ ( μ 0 ) \gamma(K_0)\le\overline{\Phi}(\mu_0) γ ( K 0 ) ≤ Φ ( μ 0 ) and γ ( K 1 ) ≥ Φ ( κ 0 ) \gamma(K_1)\ge\Phi(\kappa_0) γ ( K 1 ) ≥ Φ ( κ 0 ) , where μ 0 , κ 0 ≥ 0 \mu_0,\kappa_0\ge0 μ 0 , κ 0 ≥ 0 , then for every θ ≥ 0 \theta\ge0 θ ≥ 0 ,
γ ( K − θ ) ≤ Φ ‾ ( μ 0 + θ ( μ 0 + κ 0 ) ) . \gamma(K_{-\theta})\le\overline{\Phi}\left(\mu_0+\theta(\mu_0+\kappa_0)\right). γ ( K − θ ) ≤ Φ ( μ 0 + θ ( μ 0 + κ 0 ) ) . Proof. By definition, 1 1 + θ K − θ + θ 1 + θ K 1 ⊆ K 0 \frac{1}{1+\theta}K_{-\theta}+\frac{\theta}{1+\theta}K_1\subseteq K_0 1 + θ 1 K − θ + 1 + θ θ K 1 ⊆ K 0 . Applying Theorem 28 with λ = θ 1 + θ \lambda=\frac{\theta}{1+\theta} λ = 1 + θ θ , K = K − θ K=K_{-\theta} K = K − θ and K ′ = K 1 K'=K_1 K ′ = K 1 , we have
1 1 + θ Φ − 1 γ ( K − θ ) + θ 1 + θ Φ − 1 γ ( K 1 ) ≤ Φ − 1 γ ( 1 1 + θ K − θ + θ 1 + θ K 1 ) \frac{1}{1+\theta}\Phi^{-1}\gamma(K_{-\theta})+\frac{\theta}{1+\theta}\Phi^{-1}\gamma(K_1)
\le\Phi^{-1}\gamma\left(\frac{1}{1+\theta}K_{-\theta}+\frac{\theta}{1+\theta}K_1\right) 1 + θ 1 Φ − 1 γ ( K − θ ) + 1 + θ θ Φ − 1 γ ( K 1 ) ≤ Φ − 1 γ ( 1 + θ 1 K − θ + 1 + θ θ K 1 ) ≤ Φ − 1 γ ( K 0 ) . ( ∵ Φ − 1 is monotone ) \le\Phi^{-1}\gamma(K_0). \qquad(\because\ \Phi^{-1}\text{ is monotone}) ≤ Φ − 1 γ ( K 0 ) . ( ∵ Φ − 1 is monotone ) By rearranging,
Φ − 1 γ ( K − θ ) ≤ ( 1 + θ ) Φ − 1 γ ( K 0 ) − θ Φ − 1 γ ( K 1 ) \Phi^{-1}\gamma(K_{-\theta})\le(1+\theta)\Phi^{-1}\gamma(K_0)-\theta\Phi^{-1}\gamma(K_1) Φ − 1 γ ( K − θ ) ≤ ( 1 + θ ) Φ − 1 γ ( K 0 ) − θ Φ − 1 γ ( K 1 ) ≤ − ( 1 + θ ) μ 0 − θ κ 0 . ( ∵ γ ( K 0 ) ≤ Φ ‾ ( μ 0 ) = Φ ( − μ 0 ) , γ ( K 1 ) ≥ Φ ( κ 0 ) ) \le-(1+\theta)\mu_0-\theta\kappa_0. \qquad(\because\ \gamma(K_0)\le\overline{\Phi}(\mu_0)=\Phi(-\mu_0),\ \gamma(K_1)\ge\Phi(\kappa_0)) ≤ − ( 1 + θ ) μ 0 − θ κ 0 . ( ∵ γ ( K 0 ) ≤ Φ ( μ 0 ) = Φ ( − μ 0 ) , γ ( K 1 ) ≥ Φ ( κ 0 )) The monotonicity of Φ \Phi Φ gives γ ( K − θ ) ≤ Φ ( − μ 0 − θ ( μ 0 + κ 0 ) ) = Φ ‾ ( μ 0 + θ ( μ 0 + κ 0 ) ) \gamma(K_{-\theta})\le\Phi(-\mu_0-\theta(\mu_0+\kappa_0))=\overline{\Phi}(\mu_0+\theta(\mu_0+\kappa_0)) γ ( K − θ ) ≤ Φ ( − μ 0 − θ ( μ 0 + κ 0 )) = Φ ( μ 0 + θ ( μ 0 + κ 0 )) .
Lemma 30. Let 1 ≤ j ≤ ℓ 1 \le j \le\ell 1 ≤ j ≤ ℓ , x ∈ F j x \in F_j x ∈ F j with ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 and a , b ∈ R a,b \in\mathbb{R} a , b ∈ R , and let
K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − 1 t ∀ y ∈ N F j − 1 ( x ) } . K_0(a,b)=\left\{g:\langle g_{\perp},w_y\rangle\leq(\sqrt{3}a-b\xi_y)t-\frac{1}{t}\ \forall y\in N_{F_{j-1}}(x)\right\}. K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − t 1 ∀ y ∈ N F j − 1 ( x ) } . Then
γ ( K 0 ( a , b ) ) ≤ p F j − 1 ( x ) P [ Z 1 ∈ [ a t , a t + 1 t ] ] P [ Z 2 ∈ [ b t , b t + 1 t ] ] . \gamma(K_0(a,b))\leq\frac{p_{F_{j-1}}(x)}{\mathbb{P}[Z_1\in[at,at+\frac{1}{t}]]\mathbb{P}[Z_2\in[bt,bt+\frac{1}{t}]]}. γ ( K 0 ( a , b )) ≤ P [ Z 1 ∈ [ a t , a t + t 1 ]] P [ Z 2 ∈ [ b t , b t + t 1 ]] p F j − 1 ( x ) . where Z 1 , Z 2 ∼ N ( 0 , 1 ) Z_1,Z_2\sim N(0,1) Z 1 , Z 2 ∼ N ( 0 , 1 ) . In particular, if a ≥ 0 a\geq0 a ≥ 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\leq e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 , then
γ ( K 0 ( a , b ) ) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 . \gamma(K_0(a,b))\leq2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}. γ ( K 0 ( a , b )) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 . Proof. For the standard Gaussian vector g ∼ N ( 0 , I ) g\sim N(0,I) g ∼ N ( 0 , I ) , write g = Z 1 v x + Z 2 v x r + g ⊥ g=Z_1v_x+Z_2v_{xr}+g_{\perp} g = Z 1 v x + Z 2 v x r + g ⊥ . Here, Z 1 , Z 2 ∼ N ( 0 , 1 ) Z_1,Z_2\sim N(0,1) Z 1 , Z 2 ∼ N ( 0 , 1 ) , and Z 1 , Z 2 , g ⊥ Z_1,Z_2,g_{\perp} Z 1 , Z 2 , g ⊥ are mutually independent. The condition that x x x is selected as a local maximum is equivalent to
⟨ g , v x ⟩ > ⟨ g , v y ⟩ ( ∀ y ∈ N F j − 1 ( x ) ) ⟺ ⟨ g , v x ⟩ > ⟨ g , − 1 2 v x + 3 2 v x y ⟩ ( ∀ y ∈ N F j − 1 ( x ) ) ⟺ ⟨ g , v x ⟩ > ⟨ g , − 1 2 v x + 3 2 ( ξ y v x r + w y ) ⟩ ( ∀ y ∈ N F j − 1 ( x ) ) ⟺ Z 1 > − 1 2 Z 1 + 3 2 ( ξ y Z 2 + ⟨ g , w y ⟩ ) ( ∀ y ∈ N F j − 1 ( x ) ) ⟺ 3 Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩ ( ∀ y ∈ N F j − 1 ( x ) ) . \begin{aligned}
\langle g,v_x\rangle&>\langle g,v_y\rangle\quad(\forall y\in N_{F_{j-1}}(x))\\
&\Longleftrightarrow\langle g,v_x\rangle>\left\langle g,-\frac{1}{2}v_x+\frac{\sqrt{3}}{2}v_{xy}\right\rangle\quad(\forall y\in N_{F_{j-1}}(x))\\
&\Longleftrightarrow\langle g,v_x\rangle>\left\langle g,-\frac{1}{2}v_x+\frac{\sqrt{3}}{2}(\xi_yv_{xr}+w_y)\right\rangle\quad(\forall y\in N_{F_{j-1}}(x))\\
&\Longleftrightarrow Z_1>-\frac{1}{2}Z_1+\frac{\sqrt{3}}{2}(\xi_yZ_2+\langle g,w_y\rangle)\quad(\forall y\in N_{F_{j-1}}(x))\\
&\Longleftrightarrow\sqrt{3}Z_1-\xi_yZ_2>\langle g_{\perp},w_y\rangle\quad(\forall y\in N_{F_{j-1}}(x)).
\end{aligned} ⟨ g , v x ⟩ > ⟨ g , v y ⟩ ( ∀ y ∈ N F j − 1 ( x )) ⟺ ⟨ g , v x ⟩ > ⟨ g , − 2 1 v x + 2 3 v x y ⟩ ( ∀ y ∈ N F j − 1 ( x )) ⟺ ⟨ g , v x ⟩ > ⟨ g , − 2 1 v x + 2 3 ( ξ y v x r + w y ) ⟩ ( ∀ y ∈ N F j − 1 ( x )) ⟺ Z 1 > − 2 1 Z 1 + 2 3 ( ξ y Z 2 + ⟨ g , w y ⟩) ( ∀ y ∈ N F j − 1 ( x )) ⟺ 3 Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩ ( ∀ y ∈ N F j − 1 ( x )) . We focus on the event ( Z 1 , Z 2 ) ∈ [ a t , a t + 1 t ] × [ b t , b t + 1 t ] (Z_1,Z_2)\in[at,at+\frac{1}{t}]\times[bt,bt+\frac{1}{t}] ( Z 1 , Z 2 ) ∈ [ a t , a t + t 1 ] × [ b t , b t + t 1 ] . If the event happens and g ∈ K 0 ( a , b ) g\in K_0(a,b) g ∈ K 0 ( a , b ) , then
⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − 1 t ≤ 3 Z 1 − ξ y Z 2 . \langle g_{\perp},w_y\rangle\leq(\sqrt{3}a-b\xi_y)t-\frac{1}{t}\leq\sqrt{3}Z_1-\xi_yZ_2. ⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − t 1 ≤ 3 Z 1 − ξ y Z 2 . Since K 0 ( a , b ) K_0(a,b) K 0 ( a , b ) depends only on g ⊥ g_{\perp} g ⊥ and Z 1 , Z 2 , g ⊥ Z_1,Z_2,g_{\perp} Z 1 , Z 2 , g ⊥ are all independent,
P [ Z 1 ∈ [ a t , a t + 1 t ] ] P [ Z 2 ∈ [ b t , b t + 1 t ] ] γ ( K 0 ( a , b ) ) = P g [ ( Z 1 , Z 2 ) ∈ [ a t , a t + 1 t ] × [ b t , b t + 1 t ] and g ∈ K 0 ( a , b ) ] ≤ P g [ 3 Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩ ( ∀ y ∈ N F j − 1 ( x ) ) ] = P g [ x is selected as a local maximum ] ≤ p F j − 1 ( x ) . \begin{aligned}
\mathbb{P}[Z_1\in[at,at+\tfrac{1}{t}]]\mathbb{P}[Z_2\in[bt,bt+\tfrac{1}{t}]]\gamma(K_0(a,b))
&=\mathbb{P}_g[(Z_1,Z_2)\in[at,at+\tfrac{1}{t}]\times[bt,bt+\tfrac{1}{t}]\text{ and }g\in K_0(a,b)]\\
&\leq\mathbb{P}_g[\sqrt{3}Z_1-\xi_yZ_2>\langle g_{\perp},w_y\rangle\ (\forall y\in N_{F_{j-1}}(x))]\\
&=\mathbb{P}_g[x\text{ is selected as a local maximum}]\\
&\leq p_{F_{j-1}}(x).
\end{aligned} P [ Z 1 ∈ [ a t , a t + t 1 ]] P [ Z 2 ∈ [ b t , b t + t 1 ]] γ ( K 0 ( a , b )) = P g [( Z 1 , Z 2 ) ∈ [ a t , a t + t 1 ] × [ b t , b t + t 1 ] and g ∈ K 0 ( a , b )] ≤ P g [ 3 Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩ ( ∀ y ∈ N F j − 1 ( x ))] = P g [ x is selected as a local maximum ] ≤ p F j − 1 ( x ) . Dividing by P [ Z 1 ∈ [ a t , a t + 1 t ] ] P [ Z 2 ∈ [ b t , b t + 1 t ] ] > 0 \mathbb{P}[Z_1\in[at,at+\frac{1}{t}]]\mathbb{P}[Z_2\in[bt,bt+\frac{1}{t}]]>0 P [ Z 1 ∈ [ a t , a t + t 1 ]] P [ Z 2 ∈ [ b t , b t + t 1 ]] > 0 gives the claim. If moreover a ≥ 0 a\geq0 a ≥ 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\leq e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 , then by the bound above,
γ ( K 0 ( a , b ) ) ≤ e − t 2 / 2 1 t 2 ϕ ( a t + 1 t ) ϕ ( ∣ b ∣ t + 1 t ) ( ∵ p F j − 1 ( x ) ≤ e − t 2 / 2 , x r e f ‘ T h e o r e m 14 ‘ , a ≥ 0 ) = 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 . \begin{aligned}
\gamma(K_0(a,b))&\leq\frac{e^{-t^2/2}}{\frac{1}{t^2}\phi(at+\frac{1}{t})\phi(|b|t+\frac{1}{t})}
\quad(\because p_{F_{j-1}}(x)\leq e^{-t^2/2},\ {xref}`Theorem 14`,\ a\geq0)\\
&=2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}.
\end{aligned} γ ( K 0 ( a , b )) ≤ t 2 1 ϕ ( a t + t 1 ) ϕ ( ∣ b ∣ t + t 1 ) e − t 2 /2 ( ∵ p F j − 1 ( x ) ≤ e − t 2 /2 , x r e f ‘ T h eor e m 14‘ , a ≥ 0 ) = 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 .
The next lemma chooses the parameters so that Ehrhard extrapolation provides both the required probability bound and the target threshold ψ \psi ψ .
Lemma 31. Let 0 < L < H 0<L<\sqrt{H} 0 < L < H , ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] and s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) . For sufficiently large n n n , there exist a 0 ≥ 3 2 a_0\ge\frac{\sqrt{3}}{2} a 0 ≥ 2 3 , κ ≥ 0 \kappa\ge0 κ ≥ 0 , 0 ≤ θ ≤ 20 f ( β ) 0\le\theta\le20f(\beta) 0 ≤ θ ≤ 20 f ( β ) , ∣ b ∣ ≤ 1 2 |b|\le\frac{1}{2} ∣ b ∣ ≤ 2 1 and 0 ≤ μ ≤ 1 2 0\le\mu\le\frac{1}{2} 0 ≤ μ ≤ 2 1 such that
a 0 2 + b 2 + μ 2 = 1 , a_0^2+b^2+\mu^2=1, a 0 2 + b 2 + μ 2 = 1 , μ + θ ( μ + κ ) = H − L 2 , \mu+\theta(\mu+\kappa)=\sqrt{H-L^2}, μ + θ ( μ + κ ) = H − L 2 , and
( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ ( s ) . (1+\theta)\sqrt{3}a_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\psi(s). ( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ ( s ) .
Proof. Since s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) , we have L 2 ρ 2 + ( H − L 2 ) = s 2 L^2\rho^2+(H-L^2)=s^2 L 2 ρ 2 + ( H − L 2 ) = s 2 .
First, suppose that s ≤ ν s\le\nu s ≤ ν . Choose
a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , κ = θ = 0. a_0=\sqrt{1-s^2},\qquad b=L\rho,\qquad\mu=\sqrt{H-L^2},\qquad\kappa=\theta=0. a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , κ = θ = 0.
By L 2 ρ 2 + ( H − L 2 ) = s 2 L^2\rho^2+(H-L^2)=s^2 L 2 ρ 2 + ( H − L 2 ) = s 2 , we have a 0 2 + b 2 + μ 2 = 1 a_0^2+b^2+\mu^2=1 a 0 2 + b 2 + μ 2 = 1 . Moreover,
∣ b ∣ , μ ≤ s ≤ ν < 1 2 , a 0 ≥ 1 − ν 2 ≥ 3 2 . |b|,\mu\le s\le\nu<\frac{1}{2},\qquad a_0\ge\sqrt{1-\nu^2}\ge\frac{\sqrt{3}}{2}. ∣ b ∣ , μ ≤ s ≤ ν < 2 1 , a 0 ≥ 1 − ν 2 ≥ 2 3 .
Also, μ + θ ( μ + κ ) = μ = H − L 2 \mu+\theta(\mu+\kappa)=\mu=\sqrt{H-L^2} μ + θ ( μ + κ ) = μ = H − L 2 .
Since θ = 0 \theta=0 θ = 0 and b = L ρ b=L\rho b = L ρ ,
( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 3 1 − s 2 = ψ ( s ) . (1+\theta)\sqrt{3}a_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\sqrt{3}\sqrt{1-s^2}=\psi(s). ( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 3 1 − s 2 = ψ ( s ) .
Now suppose that s ≥ ν s\ge\nu s ≥ ν . Choose
a 0 = 1 − ν 2 , b = ν L ρ s , μ = ν H − L 2 s , a_0=\sqrt{1-\nu^2},\qquad b=\frac{\nu L\rho}{s},\qquad\mu=\frac{\nu\sqrt{H-L^2}}{s}, a 0 = 1 − ν 2 , b = s ν L ρ , μ = s ν H − L 2 ,
κ = D μ , θ = s / ν − 1 D + 1 . \kappa=D\mu,\qquad\theta=\frac{s/\nu-1}{D+1}. κ = D μ , θ = D + 1 s / ν − 1 .
Then a 0 2 + b 2 + μ 2 = 1 a_0^2+b^2+\mu^2=1 a 0 2 + b 2 + μ 2 = 1 and μ + θ ( μ + κ ) = H − L 2 \mu+\theta(\mu+\kappa)=\sqrt{H-L^2} μ + θ ( μ + κ ) = H − L 2 . Since ν 2 = D − 3 4 D < 1 4 \nu^2=\frac{D-3}{4D}<\frac{1}{4} ν 2 = 4 D D − 3 < 4 1 , we have ∣ b ∣ , μ ≤ ν < 1 2 |b|,\mu\le\nu<\frac{1}{2} ∣ b ∣ , μ ≤ ν < 2 1 and a 0 ≥ 3 2 a_0\ge\frac{\sqrt{3}}{2} a 0 ≥ 2 3 . Since s ≤ H s\le\sqrt{H} s ≤ H , we also have
0 ≤ θ ≤ H / ν − 1 D + 1 → 1 3 ( 1 + β ) ( 2 ( 2 + 3 β ) β ( 3 β − 1 ) − 1 ) < 20 f ( β ) , 0\le\theta\le\frac{\sqrt{H}/\nu-1}{D+1}\to\frac{1}{3(1+\beta)}\left(\frac{2(2+3\beta)}{\sqrt{\beta(3\beta-1)}}-1\right)<20f(\beta), 0 ≤ θ ≤ D + 1 H / ν − 1 → 3 ( 1 + β ) 1 ( β ( 3 β − 1 ) 2 ( 2 + 3 β ) − 1 ) < 20 f ( β ) ,
since H / ν → 2 ( 2 + 3 β ) β ( 3 β − 1 ) ≤ 10 3 f ( β ) < 20 f ( β ) \sqrt{H}/\nu\to\frac{2(2+3\beta)}{\sqrt{\beta(3\beta-1)}}\le10\sqrt{3}f(\beta)<20f(\beta) H / ν → β ( 3 β − 1 ) 2 ( 2 + 3 β ) ≤ 10 3 f ( β ) < 20 f ( β ) (using 2 + 3 β ≤ 5 2+3\beta\le5 2 + 3 β ≤ 5 and β > 1 / 3 \beta>1/3 β > 1/3 ) and D + 1 → 3 ( 1 + β ) > 1 D+1\to3(1+\beta)>1 D + 1 → 3 ( 1 + β ) > 1 . Hence θ ≤ 20 f ( β ) \theta\le20f(\beta) θ ≤ 20 f ( β ) for sufficiently large n n n .
Since L ρ = ( 1 + ( D + 1 ) θ ) b L\rho=(1+(D+1)\theta)b L ρ = ( 1 + ( D + 1 ) θ ) b , κ = D μ \kappa=D\mu κ = D μ , and b 2 + μ 2 = ν 2 b^2+\mu^2=\nu^2 b 2 + μ 2 = ν 2 ,
( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = θ 2 D ( D + 1 ) . \sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\frac{\theta}{2}\sqrt{D(D+1)}. ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 2 θ D ( D + 1 ) .
Also, since a 0 = 1 − ν 2 a_0=\sqrt{1-\nu^2} a 0 = 1 − ν 2 ,
( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 1 2 D + 1 D ( 3 − ( D − 3 ) θ ) . (1+\theta)\sqrt{3}a_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\frac{1}{2}\sqrt{\frac{D+1}{D}}\left(3-(D-3)\theta\right). ( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 2 1 D D + 1 ( 3 − ( D − 3 ) θ ) .
By the definition of θ \theta θ and ν 2 = D − 3 4 D \nu^2 = \frac{D-3}{4D} ν 2 = 4 D D − 3 ,
3 − ( D − 3 ) θ = 4 D D + 1 ( 1 − ν s ) . 3-(D-3)\theta=\frac{4D}{D+1}(1-\nu s). 3 − ( D − 3 ) θ = D + 1 4 D ( 1 − ν s ) .
Moreover,
1 2 D + 1 D 4 D D + 1 = 3 1 − ν 2 . \frac{1}{2}\sqrt{\frac{D+1}{D}}\frac{4D}{D+1}=\frac{\sqrt{3}}{\sqrt{1-\nu^2}}. 2 1 D D + 1 D + 1 4 D = 1 − ν 2 3 .
Hence
( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 3 ( 1 − ν s ) 1 − ν 2 = ψ ( s ) . (1+\theta)\sqrt{3}a_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}
=\frac{\sqrt{3}(1-\nu s)}{\sqrt{1-\nu^2}}=\psi(s). ( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = 1 − ν 2 3 ( 1 − ν s ) = ψ ( s ) .
This completes the proof. □
Finally, we prove Theorem 26.
Proof of Theorem 26. Set s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) . Let a 0 a_0 a 0 , b b b , μ \mu μ , κ \kappa κ and θ \theta θ be as in Theorem 31. They satisfy
a 0 2 + b 2 + μ 2 = 1 , a_0^2+b^2+\mu^2=1, a 0 2 + b 2 + μ 2 = 1 ,
μ + θ ( μ + κ ) = H − L 2 , \mu+\theta(\mu+\kappa)=\sqrt{H-L^2}, μ + θ ( μ + κ ) = H − L 2 ,
and
( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ ( s ) . (1+\theta)\sqrt{3}a_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\psi(s). ( 1 + θ ) 3 a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ ( s ) .
Moreover, a 0 ≥ 3 2 a_0\ge\frac{\sqrt{3}}{2} a 0 ≥ 2 3 , ∣ b ∣ ≤ 1 2 |b|\le\frac{1}{2} ∣ b ∣ ≤ 2 1 , 0 ≤ μ ≤ 1 2 0\le\mu\le\frac{1}{2} 0 ≤ μ ≤ 2 1 , and 0 ≤ θ ≤ 20 f ( β ) 0\le\theta\le20f(\beta) 0 ≤ θ ≤ 20 f ( β ) . Set a = a 0 − 20 w a=a_0-20w a = a 0 − 20 w . Then a ≥ 0 a\ge0 a ≥ 0 for sufficiently large t t t , and 1 − a 2 − b 2 = μ 2 + 40 a 0 w − 400 w 2 ≥ μ 2 + 30 w 1-a^2-b^2=\mu^2+40a_0w-400w^2\ge\mu^2+30w 1 − a 2 − b 2 = μ 2 + 40 a 0 w − 400 w 2 ≥ μ 2 + 30 w . Consequently,
2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ≤ O ( t − 13 ) e − μ 2 t 2 / 2 ≤ Φ ‾ ( μ t ) . 2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}\le O(t^{-13})e^{-\mu^2t^2/2}\le\overline{\Phi}(\mu t). 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ≤ O ( t − 13 ) e − μ 2 t 2 /2 ≤ Φ ( μ t ) .
The last inequality follows from the lower bound of Theorem 12, Φ ‾ ( μ t ) ≥ c 0 ( μ t + 1 ) − 1 e − μ 2 t 2 / 2 \overline{\Phi}(\mu t)\ge c_0(\mu t+1)^{-1}e^{-\mu^2t^2/2} Φ ( μ t ) ≥ c 0 ( μ t + 1 ) − 1 e − μ 2 t 2 /2 , together with μ ≤ 1 2 \mu\le\frac{1}{2} μ ≤ 2 1 .
Let
K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − 1 t ∀ y ∈ N F j − 1 ( x ) } K_0(a,b)=\left\{g:\langle g_\perp,w_y\rangle\le(\sqrt{3}a-b\xi_y)t-\frac{1}{t}\ \forall y\in N_{F_{j-1}}(x)\right\} K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( 3 a − b ξ y ) t − t 1 ∀ y ∈ N F j − 1 ( x ) }
be the set in Theorem 30. Then
γ ( K 0 ( a , b ) ) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ( ∵ x r e f ‘ T h e o r e m 30 ‘ , a ≥ 0 , p F j − 1 ( x ) ≤ e − t 2 / 2 ) ≤ Φ ‾ ( μ t ) . ( ∵ the inequality above ) \begin{aligned}
\gamma(K_0(a,b))\le2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}
\qquad(\because\ {xref}`Theorem 30`,\ a\ge0,\ p_{F_{j-1}}(x)\le e^{-t^2/2}) \\
&\le\overline{\Phi}(\mu t).
\qquad(\because\ \text{the inequality above})
\end{aligned} γ ( K 0 ( a , b )) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ( ∵ x r e f ‘ T h eor e m 30‘ , a ≥ 0 , p F j − 1 ( x ) ≤ e − t 2 /2 ) ≤ Φ ( μ t ) . ( ∵ the inequality above )
Next, let
K 1 = { g : ⟨ g ⊥ , w y ⟩ ≤ D + κ 2 1 − ( ξ y ) 2 t ∀ y ∈ N F j − 1 ( x ) } . K_1=\left\{g:\langle g_\perp,w_y\rangle\le\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}\,t\ \forall y\in N_{F_{j-1}}(x)\right\}. K 1 = { g : ⟨ g ⊥ , w y ⟩ ≤ D + κ 2 1 − ( ξ y ) 2 t ∀ y ∈ N F j − 1 ( x ) } .
For a fixed y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) , since ∥ w y ∥ = 1 − ( ξ y ) 2 \lVert w_y\rVert=\sqrt{1-(\xi_y)^2} ∥ w y ∥ = 1 − ( ξ y ) 2 , we have P g [ ⟨ g , w y ⟩ > D + κ 2 1 − ( ξ y ) 2 t ] ≤ Φ ‾ ( D + κ 2 t ) \mathbb{P}_g[\langle g,w_y\rangle>\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}t]\le\overline{\Phi}(\sqrt{D+\kappa^2}t) P g [⟨ g , w y ⟩ > D + κ 2 1 − ( ξ y ) 2 t ] ≤ Φ ( D + κ 2 t ) . Since ∣ N F j − 1 ( x ) ∣ ≤ n β = e D t 2 / 2 |N_{F_{j-1}}(x)|\le n^\beta=e^{Dt^2/2} ∣ N F j − 1 ( x ) ∣ ≤ n β = e D t 2 /2 , by a union bound, 1 − γ ( K 1 ) ≤ e D t 2 / 2 Φ ‾ ( D + κ 2 t ) ≤ Φ ‾ ( κ t ) 1-\gamma(K_1)\le e^{Dt^2/2}\overline{\Phi}(\sqrt{D+\kappa^2}t)\le\overline{\Phi}(\kappa t) 1 − γ ( K 1 ) ≤ e D t 2 /2 Φ ( D + κ 2 t ) ≤ Φ ( κ t ) , where the last inequality is Theorem 13 applied with u = κ t u=\kappa t u = κ t and v = D + κ 2 t v=\sqrt{D+\kappa^2}\,t v = D + κ 2 t . Hence γ ( K 1 ) ≥ Φ ( κ t ) \gamma(K_1)\ge\Phi(\kappa t) γ ( K 1 ) ≥ Φ ( κ t ) . Since ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_\perp,w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ , the sets K 0 K_0 K 0 and K 1 K_1 K 1 in Theorem 29 for Y = N F j − 1 ( x ) Y=N_{F_{j-1}}(x) Y = N F j − 1 ( x ) , z y = w y z_y=w_y z y = w y , A y = ( 3 a − b ξ y ) t − 1 t A_y=(\sqrt{3}a-b\xi_y)t-\frac{1}{t} A y = ( 3 a − b ξ y ) t − t 1 and B y = D + κ 2 1 − ( ξ y ) 2 B_y=\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2} B y = D + κ 2 1 − ( ξ y ) 2 are K 0 ( a , b ) K_0(a,b) K 0 ( a , b ) and K 1 K_1 K 1 above. Hence Theorem 29 with μ 0 = μ t \mu_0=\mu t μ 0 = μ t and κ 0 = κ t \kappa_0=\kappa t κ 0 = κ t gives γ ( K − θ ) ≤ Φ ( ( μ + θ ( μ + κ ) ) t ) \gamma(K_{-\theta})\leq\Phi((\mu+\theta(\mu+\kappa))t) γ ( K − θ ) ≤ Φ (( μ + θ ( μ + κ )) t ) , and the complement of K − θ K_{-\theta} K − θ is the event below. Together with μ + θ ( μ + κ ) = H − L 2 \mu+\theta(\mu+\kappa)=\sqrt{H-L^2} μ + θ ( μ + κ ) = H − L 2 ,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ D + κ 2 1 − ( ξ y ) 2 t − 1 + θ t ] ≥ Φ ( H − L 2 t ) . \mathbb{P}_g\left[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle>(1+\theta)(\sqrt{3}a-b\xi_y)t-\theta\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}t-\frac{1+\theta}{t}\right]\geq\Phi\left(\sqrt{H-L^2}t\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ D + κ 2 1 − ( ξ y ) 2 t − t 1 + θ ] ≥ Φ ( H − L 2 t ) .
When the event above happens,
⟨ g , w y ⟩ + L ρ ξ y t > ( ( 1 + θ ) 3 a + ( L ρ − ( 1 + θ ) b ) ξ y − θ D + κ 2 1 − ( ξ y ) 2 ) t − 1 + θ t . \langle g,w_y\rangle+L\rho\xi_y t>\left((1+\theta)\sqrt{3}a+(L\rho-(1+\theta)b)\xi_y-\theta\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}\right)t-\frac{1+\theta}{t}. ⟨ g , w y ⟩ + L ρ ξ y t > ( ( 1 + θ ) 3 a + ( L ρ − ( 1 + θ ) b ) ξ y − θ D + κ 2 1 − ( ξ y ) 2 ) t − t 1 + θ .
Since A ξ − B 1 − ξ 2 ≥ − A 2 + B 2 A\xi-B\sqrt{1-\xi^2}\geq-\sqrt{A^2+B^2} A ξ − B 1 − ξ 2 ≥ − A 2 + B 2 for ∣ ξ ∣ ≤ 1 |\xi|\leq1 ∣ ξ ∣ ≤ 1 by Cauchy–Schwarz, we obtain
⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ( 1 + θ ) 3 a − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − 1 + θ t = ( ( 1 + θ ) 3 ( a 0 − 20 w ) − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − 1 + θ t ≥ ( ψ ( s ) − 20 3 ( 1 + 20 f ( β ) ) w ) t − 1 + 20 f ( β ) t ≥ ( ψ ( s ) − 40 ( 1 + 20 f ( β ) ) w ) t . \begin{aligned}
\langle g,w_y\rangle+L\rho\xi_y t&\geq\left((1+\theta)\sqrt{3}a-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}\right)t-\frac{1+\theta}{t}\\
&=\left((1+\theta)\sqrt{3}(a_0-20w)-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}\right)t-\frac{1+\theta}{t}\\
&\geq\left(\psi(s)-20\sqrt{3}(1+20f(\beta))w\right)t-\frac{1+20f(\beta)}{t}\\
&\geq\left(\psi(s)-40(1+20f(\beta))w\right)t.
\end{aligned} ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ( 1 + θ ) 3 a − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − t 1 + θ = ( ( 1 + θ ) 3 ( a 0 − 20 w ) − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − t 1 + θ ≥ ( ψ ( s ) − 20 3 ( 1 + 20 f ( β )) w ) t − t 1 + 20 f ( β ) ≥ ( ψ ( s ) − 40 ( 1 + 20 f ( β )) w ) t .
where the second inequality uses θ ≤ 20 f ( β ) \theta\leq20f(\beta) θ ≤ 20 f ( β ) , and the last one uses ( 40 − 20 3 ) ( 1 + 20 f ( β ) ) w t ≥ ( 1 + 20 f ( β ) ) / t (40-20\sqrt{3})(1+20f(\beta))wt\geq(1+20f(\beta))/t ( 40 − 20 3 ) ( 1 + 20 f ( β )) w t ≥ ( 1 + 20 f ( β )) / t for sufficiently large t t t . This completes the proof. □
Iterated Lifting
In this subsection, we repeatedly apply the single-step lifting lemma (Theorem 25), starting from the initial Gaussian cover around a vertex in the deepest core. After o ( ℓ ) o(\ell) o ( ℓ ) lifting steps, the cover threshold exceeds H t \sqrt{H}t H t , while the covering probability decreases by only o ( 1 ) o(1) o ( 1 ) .
Lemma 32. Suppose that every empirical estimate used in Algorithm 2 satisfies ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ / 8 |\widehat{p}_j(x)-p_{F_{j-1}}(x)|\leq\delta/8 ∣ p j ( x ) − p F j − 1 ( x ) ∣ ≤ δ /8 for every level j j j at which F j F_j F j is constructed and every x ∈ F j − 1 x\in F_{j-1} x ∈ F j − 1 . Then, for sufficiently large n n n , Algorithm 2 returns SUCCESS( I ) (I) ( I ) for some independent set I I I with ∣ I ∣ ≥ T |I|\geq T ∣ I ∣ ≥ T .
Proof. Suppose for contradiction that Algorithm 2 returns FAIL. Then every independent set sampled at every level has size less than T T T , all cores F 1 , … , F ℓ F_1,\ldots,F_\ell F 1 , … , F ℓ are constructed, and F ℓ F_\ell F ℓ is nonempty by Theorem 20. Moreover, by the definition of F j F_j F j , every x ∈ F j x\in F_j x ∈ F j satisfies p ^ j ( x ) < δ / 2 \widehat{p}_j(x)<\delta/2 p j ( x ) < δ /2 , and hence
p F j − 1 ( x ) ≤ p ^ j ( x ) + δ 8 < e − t 2 / 2 ( ∀ j , ∀ x ∈ F j ) . p_{F_{j-1}}(x)\leq\widehat{p}_j(x)+\frac{\delta}{8}<e^{-t^2/2}\qquad(\forall j,\ \forall x\in F_j). p F j − 1 ( x ) ≤ p j ( x ) + 8 δ < e − t 2 /2 ( ∀ j , ∀ x ∈ F j ) .
Choose a root r ∈ F ℓ r\in F_\ell r ∈ F ℓ , and let W 0 W_0 W 0 and U 0 U_0 U 0 be as in Subsection 4.3. Since p F ℓ − 1 ( r ) < δ = e − t 2 / 2 p_{F_{\ell-1}}(r)<\delta=e^{-t^2/2} p F ℓ − 1 ( r ) < δ = e − t 2 /2 , Theorem 24 shows that { u x } x ∈ W 0 \{u_x\}_{x\in W_0} { u x } x ∈ W 0 is a ( U 0 t , 1 − O ( t e − t 2 / 4 ) ) (U_0t,1-O(te^{-t^2/4})) ( U 0 t , 1 − O ( t e − t 2 /4 )) -cover. Let m m m be the smallest integer with U m > H U_m>\sqrt{H} U m > H . Since H = O ( 1 ) H=O(1) H = O ( 1 ) , we have m = O ( 1 / w ) = O ( t 2 / log t ) m=O(1/w)=O(t^2/\log t) m = O ( 1/ w ) = O ( t 2 / log t ) . As t 2 = Θ ( log n ) t^2=\Theta(\log n) t 2 = Θ ( log n ) and ℓ = Θ ( log n ) \ell=\Theta(\log n) ℓ = Θ ( log n ) , this gives m = o ( ℓ ) m=o(\ell) m = o ( ℓ ) , so m ≤ ℓ − 1 m\leq\ell-1 m ≤ ℓ − 1 for sufficiently large n n n . For 0 ≤ i < m 0\leq i<m 0 ≤ i < m , define W i + 1 = N F ℓ − i − 2 ( W i ) W_{i+1}=N_{F_{\ell-i-2}}(W_i) W i + 1 = N F ℓ − i − 2 ( W i ) . By construction, W i ⊆ F ℓ − i − 1 W_i\subseteq F_{\ell-i-1} W i ⊆ F ℓ − i − 1 and W i W_i W i is determined by the samples used to construct the cores, independently of g g g .
We claim that { u x } x ∈ W i \{u_x\}_{x\in W_i} { u x } x ∈ W i is a ( U i t , 1 − O ( t e − t 2 / 4 ) − i t − 18 ) (U_i t,1-O(te^{-t^2/4})-it^{-18}) ( U i t , 1 − O ( t e − t 2 /4 ) − i t − 18 ) -cover for every 0 ≤ i ≤ m 0\leq i\leq m 0 ≤ i ≤ m . The case i = 0 i=0 i = 0 is Theorem 24, as noted above. For i < m i<m i < m , every x ∈ W i ⊆ F ℓ − i − 1 x\in W_i\subseteq F_{\ell-i-1} x ∈ W i ⊆ F ℓ − i − 1 satisfies p F ℓ − i − 2 ( x ) ≤ e − t 2 / 2 p_{F_{\ell-i-2}}(x)\leq e^{-t^2/2} p F ℓ − i − 2 ( x ) ≤ e − t 2 /2 by the bound above. Since 0 < U i ≤ H 0<U_i\leq\sqrt{H} 0 < U i ≤ H , Theorem 25 applied to W = W i W=W_i W = W i and U = U i U=U_i U = U i shows that { u x } x ∈ W i + 1 \{u_x\}_{x\in W_{i+1}} { u x } x ∈ W i + 1 is a ( U i + 1 t , 1 − O ( t e − t 2 / 4 ) − ( i + 1 ) t − 18 ) (U_{i+1}t,1-O(te^{-t^2/4})-(i+1)t^{-18}) ( U i + 1 t , 1 − O ( t e − t 2 /4 ) − ( i + 1 ) t − 18 ) -cover, completing the induction.
Since m = O ( t 2 / log t ) m=O(t^2/\log t) m = O ( t 2 / log t ) , we have m t − 18 = o ( 1 ) mt^{-18}=o(1) m t − 18 = o ( 1 ) , and also t e − t 2 / 4 = o ( 1 ) te^{-t^2/4}=o(1) t e − t 2 /4 = o ( 1 ) . Thus, for sufficiently large n n n , { u x } x ∈ W m \{u_x\}_{x\in W_m} { u x } x ∈ W m is a ( U m t , 1 2 ) (U_m t,\frac{1}{2}) ( U m t , 2 1 ) -cover.
On the other hand, ∥ u x ∥ ≤ 1 \|u_x\|\leq1 ∥ u x ∥ ≤ 1 for every x x x , so P g [ ⟨ g , u x ⟩ ≥ U m t ] ≤ Φ ‾ ( U m t ) \mathbb{P}_g[\langle g,u_x\rangle\geq U_m t]\leq\overline{\Phi}(U_m t) P g [⟨ g , u x ⟩ ≥ U m t ] ≤ Φ ( U m t ) , and ∣ W m ∣ ≤ n = e H 0 t 2 / 2 |W_m|\leq n=e^{H_0t^2/2} ∣ W m ∣ ≤ n = e H 0 t 2 /2 . Therefore, by a union bound and the Gaussian tail bound Φ ‾ ( u ) ≤ e − u 2 / 2 \overline{\Phi}(u)\leq e^{-u^2/2} Φ ( u ) ≤ e − u 2 /2 of Theorem 12,
P g [ max x ∈ W m ⟨ g , u x ⟩ ≥ U m t ] ≤ n Φ ‾ ( U m t ) ≤ e H 0 t 2 / 2 e − U m 2 t 2 / 2 < e − ( H − H 0 ) t 2 / 2 = t − 20 , \begin{aligned}
\mathbb{P}_g\left[\max_{x\in W_m}\langle g,u_x\rangle\geq U_m t\right]\leq n\overline{\Phi}(U_m t) \\
&\leq e^{H_0t^2/2}e^{-U_m^2t^2/2} \\
&<e^{-(H-H_0)t^2/2}=t^{-20},
\end{aligned} P g [ x ∈ W m max ⟨ g , u x ⟩ ≥ U m t ] ≤ n Φ ( U m t ) ≤ e H 0 t 2 /2 e − U m 2 t 2 /2 < e − ( H − H 0 ) t 2 /2 = t − 20 ,
where we used U m 2 > H U_m^2>H U m 2 > H and ( H − H 0 ) t 2 / 2 = 20 log t (H-H_0)t^2/2=20\log t ( H − H 0 ) t 2 /2 = 20 log t . For sufficiently large n n n , this is less than 1 2 \frac{1}{2} 2 1 , contradicting the ( U m t , 1 2 ) (U_m t,\frac{1}{2}) ( U m t , 2 1 ) -cover property. Hence Algorithm 2 cannot return FAIL and must return SUCCESS( I ) (I) ( I ) for an independent set I I I with ∣ I ∣ ≥ T |I|\geq T ∣ I ∣ ≥ T .
Lemma 33 (High-degree regime). Fix 1 / 3 < β ≤ 1 1/3<\beta\leq1 1/3 < β ≤ 1 , and let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β , given together with a strict vector 3-coloring. Then, there is a randomized polynomial-time algorithm that, with probability at least 3 4 \frac{3}{4} 4 3 , returns an independent set I I I with ∣ I ∣ ≥ n 2 + 2 β 2 + 3 β − o ( 1 ) |I|\geq n^{\frac{2+2\beta}{2+3\beta}-o(1)} ∣ I ∣ ≥ n 2 + 3 β 2 + 2 β − o ( 1 ) .
Proof. Run one trial of Algorithm 2 with the high-degree parameters of Table 2 . Since β \beta β is fixed and ℓ = O ( log n ) \ell=O(\log n) ℓ = O ( log n ) , we have δ − 1 = n β / ( 2 + 3 β ) polylog ( n ) \delta^{-1}=n^{\beta/(2+3\beta)}\operatorname{polylog}(n) δ − 1 = n β / ( 2 + 3 β ) polylog ( n ) . Each call to Algorithm 1 takes polynomial time, and the number of calls is at most ℓ R \ell R ℓ R . Since R = O ( δ − 2 log n ) = n 2 β / ( 2 + 3 β ) + o ( 1 ) R=O(\delta^{-2}\log n)=n^{2\beta/(2+3\beta)+o(1)} R = O ( δ − 2 log n ) = n 2 β / ( 2 + 3 β ) + o ( 1 ) , one trial runs in polynomial time.
By Theorem 21, with probability at least 3 4 \frac{3}{4} 4 3 , every empirical estimate used in the construction of the cores satisfies ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ / 8 |\widehat{p}_j(x)-p_{F_{j-1}}(x)|\leq\delta/8 ∣ p j ( x ) − p F j − 1 ( x ) ∣ ≤ δ /8 . Conditioned on this event, Theorem 32 shows that Algorithm 2 returns SUCCESS( I ) (I) ( I ) for an independent set I I I with
∣ I ∣ ≥ T ≥ n δ 100 ℓ 4 = n 1 − β 2 + 3 β − o ( 1 ) = n 2 + 2 β 2 + 3 β − o ( 1 ) , |I|\geq T\geq\frac{n\delta}{100\ell^4}=n^{1-\frac{\beta}{2+3\beta}-o(1)}=n^{\frac{2+2\beta}{2+3\beta}-o(1)}, ∣ I ∣ ≥ T ≥ 100 ℓ 4 n δ = n 1 − 2 + 3 β β − o ( 1 ) = n 2 + 3 β 2 + 2 β − o ( 1 ) ,
when n n n is sufficiently large. The “− o ( 1 ) -o(1) − o ( 1 ) ” part allows us to ignore the case of small n n n , and Theorem 33 follows.
Theorem 3. Fix β ∈ ( 0 , 1 ] \beta\in(0,1] β ∈ ( 0 , 1 ] . Let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β , given together with a strict vector 3-coloring, that is, unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V with ⟨ v x , v y ⟩ = − 1 2 \langle v_x,v_y\rangle=-\frac{1}{2} ⟨ v x , v y ⟩ = − 2 1 for every edge x y ∈ E xy\in E x y ∈ E . For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , returns an independent set I I I with
∣ I ∣ ≥ { n 1 − β / 3 − o ( 1 ) , 0 < β ≤ 1 / 3 , n 2 + 2 β 2 + 3 β − o ( 1 ) , 1 / 3 < β ≤ 1. |I|\geq
\begin{cases}
n^{1-\beta/3-o(1)}, & 0<\beta\leq1/3,\\
n^{\frac{2+2\beta}{2+3\beta}-o(1)}, & 1/3<\beta\leq1.
\end{cases} ∣ I ∣ ≥ { n 1 − β /3 − o ( 1 ) , n 2 + 3 β 2 + 2 β − o ( 1 ) , 0 < β ≤ 1/3 , 1/3 < β ≤ 1.
Proof of Theorem 3. Fix C > 0 C > 0 C > 0 . Run ⌈ C log 4 n ⌉ \lceil C\log_4 n\rceil ⌈ C log 4 n ⌉ independent trials of Algorithm 2, using the parameters from Theorem 18 if 0 < β ≤ 1 / 3 0 < \beta\le1/3 0 < β ≤ 1/3 and from Theorem 33 if 1 / 3 < β ≤ 1 1/3 < \beta\le1 1/3 < β ≤ 1 . Return the largest independent set obtained. By Theorem 18 or Theorem 33 , according to the regime, each trial returns an independent set of the required size with probability at least 3 4 \frac{3}{4} 4 3 . Hence the probability that no trial returns such a set is at most 4 − ⌈ C log 4 n ⌉ ≤ n − C 4^{-\lceil C\log_4 n\rceil} \le n^{-C} 4 − ⌈ C l o g 4 n ⌉ ≤ n − C . Since each trial runs in polynomial time and there are O ( log n ) O(\log n) O ( log n ) trials, the total running time remains polynomial. □ \square □
Improved Sparse Progress
In this section, we prove Theorem 5 using the preceding vector coloring algorithm as a subroutine. Table 3 lists the parameters used in this section.
Table 3. Parameters used in § 5 .
As in § 4 , U 0 = 3 / ( 2 2 ) U_0=3/(2\sqrt{2}) U 0 = 3/ ( 2 2 ) is the initial Gaussian cover threshold divided by t t t . We write A 0 = D a + 1 4 A_0=\sqrt{D_a+\frac{1}{4}} A 0 = D a + 4 1 for the initial tail-bound threshold divided by t t t .
Preliminaries for SparseProgress
We consider the following semidefinite programming relaxation: for m ∈ { 1 , 2 } m\in\{1,2\} m ∈ { 1 , 2 } and S ⊆ V S\subseteq V S ⊆ V , let Θ m ( S ) \Theta_m(S) Θ m ( S ) be the degree-2 relaxation of the problem of finding a largest induced subgraph of G [ S ] G[S] G [ S ] that is m m m -colorable. For each vertex i ∈ S i\in S i ∈ S and color c ∈ { 1 , … , m } c\in\{1,\ldots,m\} c ∈ { 1 , … , m } we have a vector x i , c x_{i,c} x i , c , together with an anchor vector x ∅ x_{\emptyset} x ∅ , and we write
x i = ∑ c = 1 m x i , c . x_i = \sum_{c=1}^{m} x_{i,c}. x i = c = 1 ∑ m x i , c .
Then
Θ m ( S ) = max ∑ i ∈ S ∑ c = 1 m ∥ x i , c ∥ 2 \Theta_m(S) = \max\sum_{i \in S} \sum_{c=1}^{m} \lVert x_{i,c} \rVert^2 Θ m ( S ) = max i ∈ S ∑ c = 1 ∑ m ∥ x i , c ∥ 2
subject to
∥ x ∅ ∥ 2 = 1 , \lVert x_{\emptyset} \rVert^2 = 1, ∥ x ∅ ∥ 2 = 1 ,
⟨ x ∅ , x i , c ⟩ = ∥ x i , c ∥ 2 for all i ∈ S and c ∈ { 1 , … , m } , \langle x_{\emptyset}, x_{i,c} \rangle= \lVert x_{i,c} \rVert^2 \qquad\text{for all } i \in S \text{ and } c \in\{1,\ldots,m\}, ⟨ x ∅ , x i , c ⟩ = ∥ x i , c ∥ 2 for all i ∈ S and c ∈ { 1 , … , m } ,
⟨ x i , c , x i , c ′ ⟩ = 0 for all i ∈ S and c ≠ c ′ ∈ { 1 , … , m } , \langle x_{i,c}, x_{i,c'} \rangle= 0 \qquad\text{for all } i \in S \text{ and } c \ne c' \in\{1,\ldots,m\}, ⟨ x i , c , x i , c ′ ⟩ = 0 for all i ∈ S and c = c ′ ∈ { 1 , … , m } ,
⟨ x i , c , x j , c ⟩ = 0 for all i j ∈ E [ S ] and c ∈ { 1 , … , m } , \langle x_{i,c}, x_{j,c} \rangle= 0 \qquad\text{for all } ij \in E[S] \text{ and } c \in\{1,\ldots,m\}, ⟨ x i , c , x j , c ⟩ = 0 for all ij ∈ E [ S ] and c ∈ { 1 , … , m } ,
⟨ x ∅ − x i , x ∅ − x j ⟩ ≥ 0 for all i j ∈ E [ S ] . \langle x_{\emptyset} - x_i, x_{\emptyset} - x_j \rangle\ge0 \qquad\text{for all } ij \in E[S]. ⟨ x ∅ − x i , x ∅ − x j ⟩ ≥ 0 for all ij ∈ E [ S ] .
In the intended solution, for an induced subgraph of G [ S ] G[S] G [ S ] with a proper m m m -coloring, we set x i , c = x ∅ x_{i,c} = x_{\emptyset} x i , c = x ∅ if vertex i i i belongs to the subgraph and receives color c c c , and x i , c = 0 x_{i,c} = 0 x i , c = 0 otherwise. For m = 1 m = 1 m = 1 , without the last family of constraints, this is the Lovász theta function [44 ] . Due to the third equality, we have ∥ x i ∥ 2 = ∑ c = 1 m ∥ x i , c ∥ 2 \lVert x_i \rVert^2 = \sum_{c=1}^{m} \lVert x_{i,c} \rVert^2 ∥ x i ∥ 2 = ∑ c = 1 m ∥ x i , c ∥ 2 .
The Algorithm SparseProgress
In this subsection, we describe our improved algorithm for finding a large independent set. The algorithm uses a vector coloring SDP strengthened with additional constraints. We begin by defining the correlation bands used in the algorithm.
Bands. Fix a partition P \mathcal{P} P of [ − 1 2 , 1 ] [-\frac{1}{2},1] [ − 2 1 , 1 ] into O ( log n ) O(\log n) O ( log n ) intervals of length at most 1 / log n 1/\log n 1/ log n , each contained in [ − 1 2 , 0 ] [-\frac{1}{2},0] [ − 2 1 , 0 ] or in [ 0 , 1 ] [0,1] [ 0 , 1 ] . For a vertex r ∈ V r \in V r ∈ V and P ∈ P P \in\mathcal{P} P ∈ P , let S r , P = { x ∈ V : ⟨ v r , v x ⟩ ∈ P } S_{r,P} = \{x \in V : \langle v_r, v_x \rangle\in P\} S r , P = { x ∈ V : ⟨ v r , v x ⟩ ∈ P } . We call P P P , and the band S r , P S_{r,P} S r , P , positive if P ⊆ [ 0 , 1 ] P \subseteq[0,1] P ⊆ [ 0 , 1 ] and negative otherwise; we set ρ P = inf P \rho_P = \inf P ρ P = inf P for a positive band and ρ P = sup P \rho_P = \sup P ρ P = sup P for a negative band.
The choice of auxiliary SDPs is motivated by the simplex representation of a proper 3-coloring (Subsection 3.1). In this representation, 1 + 2 ⟨ v r , v x ⟩ 3 \frac{1+2\langle v_r,v_x\rangle}{3} 3 1 + 2 ⟨ v r , v x ⟩ is 1 if x x x has the same color as r r r and 0 otherwise, whereas 2 − 2 ⟨ v r , v x ⟩ 3 \frac{2-2\langle v_r,v_x\rangle}{3} 3 2 − 2 ⟨ v r , v x ⟩ is 1 if x x x has one of the other two colors and 0 otherwise. The vertices of the first kind form an independent set, and those of the second kind induce a bipartite subgraph. This motivates using Θ 1 \Theta_1 Θ 1 for positive bands and Θ 2 \Theta_2 Θ 2 for negative bands.
For each band, we use an auxiliary SDP witness either to extract an independent set directly or to construct an ordinary vector k k k -coloring, to which we apply the algorithm of Theorem 2 . If the resulting independent set has size at least T T T , we return it. Otherwise, provided that the internal vector coloring algorithm meets its guarantee, we obtain an upper bound on the size of the band. These bounds are then used to strengthen the tail bound on the neighborhood maximum and the Gaussian cover lifting argument.
We use BuildVector to obtain two types of guarantees. First, for each correlation band, it provides an auxiliary SDP witness whose total mass is sufficiently large. These witnesses are used in the band-rounding procedures. Second, the strengthened SDP imposes additional constraints that control correlations between Gaussian projections associated with adjacent vertices. These constraints will be used in the tail analysis of the neighborhood maximum in Subsection 5.4.
Certified vectors and band witnesses. Let ε S D P \varepsilon_{\mathrm{SDP}} ε SDP be the accuracy parameter of Subsection 3.1. To motivate the additional semidefinite constraint, we first assume that the vector coloring constraints are exact. For an edge r x ∈ E rx \in E r x ∈ E , consider the symmetric matrix
Γ r x ′ = ( ( ⟨ v r y , v r y ′ ⟩ ) y , y ′ ( 2 ⟨ v r y , v x z ⟩ ) y , z ( 2 ⟨ v x z , v r y ⟩ ) z , y ( ⟨ v x z , v x z ′ ⟩ ) z , z ′ ) , \Gamma'_{rx} =
\begin{pmatrix}
(\langle v_{ry},v_{ry'}\rangle)_{y,y'} & (2\langle v_{ry},v_{xz}\rangle)_{y,z} \\
(2\langle v_{xz},v_{ry}\rangle)_{z,y} & (\langle v_{xz},v_{xz'}\rangle)_{z,z'}
\end{pmatrix}, Γ r x ′ = ( (⟨ v r y , v r y ′ ⟩ ) y , y ′ ( 2 ⟨ v x z , v r y ⟩ ) z , y ( 2 ⟨ v r y , v x z ⟩ ) y , z (⟨ v x z , v x z ′ ⟩ ) z , z ′ ) ,
where y , y ′ ∈ N V ( r ) y,y' \in N_V(r) y , y ′ ∈ N V ( r ) and z , z ′ ∈ N V ( x ) z,z' \in N_V(x) z , z ′ ∈ N V ( x ) .
Let X ∈ R V × V X \in\mathbb{R}^{V \times V} X ∈ R V × V be a symmetric matrix with unit diagonal and define
( Γ r x ( X ) ) ( i , j ) , ( i ′ , j ′ ) = 1 + 1 i ≠ i ′ 3 ( 4 X j j ′ + 2 X j i ′ + 2 X i j ′ + X i i ′ ) \left(\Gamma_{rx}(X)\right)_{(i,j),(i',j')} = \frac{1+\mathbf{1}_{i\ne i'}}{3}\left(4X_{jj'}+2X_{ji'}+2X_{ij'}+X_{ii'}\right) ( Γ r x ( X ) ) ( i , j ) , ( i ′ , j ′ ) = 3 1 + 1 i = i ′ ( 4 X j j ′ + 2 X j i ′ + 2 X i j ′ + X i i ′ )
for i , i ′ ∈ { r , x } i,i' \in\{r,x\} i , i ′ ∈ { r , x } , j ∈ N V ( i ) j \in N_V(i) j ∈ N V ( i ) , and j ′ ∈ N V ( i ′ ) j' \in N_V(i') j ′ ∈ N V ( i ′ ) . If X X X is the Gram matrix of { v i } i ∈ V \{v_i\}_{i\in V} { v i } i ∈ V , then Γ r x ( X ) = Γ r x ′ \Gamma_{rx}(X)=\Gamma'_{rx} Γ r x ( X ) = Γ r x ′ , since v i j = ( 2 v j + v i ) / 3 v_{ij}=(2v_j+v_i)/\sqrt{3} v ij = ( 2 v j + v i ) / 3 for every i j ∈ E ij\in E ij ∈ E . In the exact setting, positive semidefiniteness of this matrix bounds the correlation between two spans of edge residuals by 1 2 \frac{1}{2} 2 1 , as shown in Theorem 41 .
We use BuildVector through the following output guarantee.
Lemma 34 (Certified vectors and band witnesses). There is a subroutine BuildVector that, given a 3-colorable graph G = ( V , E ) G=(V,E) G = ( V , E ) on n ≥ 2 n\ge2 n ≥ 2 vertices, returns ( { v i } i ∈ V , S = { ( ρ P , S r , P , W r , P ) } r ∈ V , P ∈ P ) \left(\{v_i\}_{i\in V},\mathcal{S}=\{(\rho_P,S_{r,P},\mathcal{W}_{r,P})\}_{r\in V,P\in\mathcal{P}}\right) ( { v i } i ∈ V , S = {( ρ P , S r , P , W r , P ) } r ∈ V , P ∈ P ) in polynomial time in n n n and log ( 1 / ε S D P ) \log(1/\varepsilon_{\mathrm{SDP}}) log ( 1/ ε SDP ) and with the following properties.
The vectors { v i } i ∈ V \{v_i\}_{i\in V} { v i } i ∈ V are unit vectors satisfying
− 1 2 ≤ ⟨ v i , v j ⟩ ≤ 1 for all i , j ∈ V , -\frac{1}{2}\le\langle v_i,v_j\rangle\le1 \qquad\text{for all }i,j\in V, − 2 1 ≤ ⟨ v i , v j ⟩ ≤ 1 for all i , j ∈ V , ⟨ v i , v j ⟩ ≤ − 1 2 + ε S D P for all i j ∈ E , \langle v_i,v_j\rangle\le-\frac{1}{2}+\varepsilon_{\mathrm{SDP}} \qquad\text{for all }ij\in E, ⟨ v i , v j ⟩ ≤ − 2 1 + ε SDP for all ij ∈ E , ⟨ v i , v j ⟩ + ⟨ v i , v j ′ ⟩ ≤ 1 2 for all j j ′ ∈ E and i ∈ V ∖ { j , j ′ } , \langle v_i,v_j\rangle+\langle v_i,v_{j'}\rangle\le\frac{1}{2} \qquad\text{for all }jj'\in E\text{ and }i\in V\setminus\{j,j'\}, ⟨ v i , v j ⟩ + ⟨ v i , v j ′ ⟩ ≤ 2 1 for all j j ′ ∈ E and i ∈ V ∖ { j , j ′ } , Γ r x ( X ) ⪰ − ε S D P I for all r x ∈ E , \Gamma_{rx}(X)\succeq-\varepsilon_{\mathrm{SDP}}I \qquad\text{for all }rx\in E, Γ r x ( X ) ⪰ − ε SDP I for all r x ∈ E , where X i j = ⟨ v i , v j ⟩ X_{ij}=\langle v_i,v_j\rangle X ij = ⟨ v i , v j ⟩ .
For every r ∈ V r\in V r ∈ V and P ∈ P P\in\mathcal{P} P ∈ P , writing S = S r , P S=S_{r,P} S = S r , P , the triple ( ρ P , S r , P , W r , P ) (\rho_P,S_{r,P},\mathcal{W}_{r,P}) ( ρ P , S r , P , W r , P ) has the following properties.
(i) If P P P is positive, then W r , P = ( x ∅ , { x j } j ∈ S ) \mathcal{W}_{r,P}=(x_{\varnothing},\{x_j\}_{j\in S}) W r , P = ( x ∅ , { x j } j ∈ S ) is feasible for Θ 1 ( S ) \Theta_1(S) Θ 1 ( S ) and
∑ j ∈ S ∥ x j ∥ 2 ≥ ∑ j ∈ S 1 + 2 ⟨ v r , v j ⟩ 3 − ε S D P ≥ 1 + 2 ρ P 3 ∣ S ∣ − ε S D P . \sum_{j\in S}\|x_j\|^2 \ge\sum_{j\in S}\frac{1+2\langle v_r,v_j\rangle}{3}-\varepsilon_{\mathrm{SDP}}\ge\frac{1+2\rho_P}{3}|S|-\varepsilon_{\mathrm{SDP}}. j ∈ S ∑ ∥ x j ∥ 2 ≥ j ∈ S ∑ 3 1 + 2 ⟨ v r , v j ⟩ − ε SDP ≥ 3 1 + 2 ρ P ∣ S ∣ − ε SDP . (ii) If P P P is negative, then W r , P = ( x ∅ , { ( x j , 1 , x j , 2 ) } j ∈ S ) \mathcal{W}_{r,P}=(x_{\varnothing},\{(x_{j,1},x_{j,2})\}_{j\in S}) W r , P = ( x ∅ , {( x j , 1 , x j , 2 ) } j ∈ S ) is feasible for Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) and
∑ j ∈ S ( ∥ x j , 1 ∥ 2 + ∥ x j , 2 ∥ 2 ) ≥ ∑ j ∈ S 2 − 2 ⟨ v r , v j ⟩ 3 − ε S D P ≥ 2 − 2 ρ P 3 ∣ S ∣ − ε S D P . \sum_{j\in S}\left(\|x_{j,1}\|^2+\|x_{j,2}\|^2\right)\ge\sum_{j\in S}\frac{2-2\langle v_r,v_j\rangle}{3}-\varepsilon_{\mathrm{SDP}}\ge\frac{2-2\rho_P}{3}|S|-\varepsilon_{\mathrm{SDP}}. j ∈ S ∑ ( ∥ x j , 1 ∥ 2 + ∥ x j , 2 ∥ 2 ) ≥ j ∈ S ∑ 3 2 − 2 ⟨ v r , v j ⟩ − ε SDP ≥ 3 2 − 2 ρ P ∣ S ∣ − ε SDP .
The implementation of BuildVector and the proof of Theorem 34 are given in Appendix § 9 . We use the algorithm of Theorem 2 as a subroutine on the vector colorings obtained from the bands. We set a rounding slack ε t h r ∈ ( 0 , 1 / 30 ] \varepsilon_{\mathrm{thr}} \in(0,1/30] ε thr ∈ ( 0 , 1/30 ] for Algorithm 4 and Algorithm 5, and fix a grid step h ∈ ( 0 , 1 / 8 ] h \in(0,1/8] h ∈ ( 0 , 1/8 ] with 1 / h ∈ N 1/h \in\mathbb{N} 1/ h ∈ N for VectorColoring. These parameters are determined so that they are sufficiently small depending only on β \beta β and η \eta η of Table 3 , and not on n n n . In particular, they are chosen so that the term O ( h + ε t h r ) O(h+\varepsilon_{\mathrm{thr}}) O ( h + ε thr ) in the proof of Theorem 37, whose implied constant depends only on β \beta β , is at most η / 4 \eta/4 η /4 . Since the vector coloring parameter and the degree exponent may vary across bands, we discretize them to ensure that the asymptotic guarantee of Theorem 2 applies for all choices of these parameters. VectorColoring( G ′ , { z i } i ∈ V ( G ′ ) , k ) (G',\{z_i\}_{i\in V(G')},k) ( G ′ , { z i } i ∈ V ( G ′ ) , k ) runs the algorithm of Theorem 2 with k h = min { k ′ ∈ K h : k ′ ≥ k } k_h=\min\{k'\in\mathcal{K}_h:k'\ge k\} k h = min { k ′ ∈ K h : k ′ ≥ k } in place of k k k and β h ( G ′ ) = min { β ′ ∈ D h : Δ ( G ′ ) ≤ ∣ V ( G ′ ) ∣ β ′ } \beta_h(G')=\min\{\beta'\in\mathcal{D}_h:\Delta(G')\le|V(G')|^{\beta'}\} β h ( G ′ ) = min { β ′ ∈ D h : Δ ( G ′ ) ≤ ∣ V ( G ′ ) ∣ β ′ } in place of β \beta β , where K h = { 2 + h , 2 + 2 h , … , 4 } \mathcal{K}_h=\{2+h,2+2h,\ldots,4\} K h = { 2 + h , 2 + 2 h , … , 4 } , D h = { h , 2 h , … , 1 } \mathcal{D}_h=\{h,2h,\ldots,1\} D h = { h , 2 h , … , 1 } , and h h h is the grid step fixed above. Each invocation uses the size and maximum degree of its own input graph G ′ G' G ′ , rather than those of the original graph.
Algorithm 3 SparseProgress
Require: graph G = ( V , E ) G=(V,E) G = ( V , E )
Ensure: an independent set I I I with ∣ I ∣ ≥ T |I|\ge T ∣ I ∣ ≥ T , or FAIL
( { v x } x ∈ V , S ) ← ({\{v_x\}}_{x\in V},\mathcal{S})\gets ( { v x } x ∈ V , S ) ← BuildVector( G ) (G) ( G )
for ( ρ P , S , W ) ∈ S (\rho_P,S,\mathcal{W})\in\mathcal{S} ( ρ P , S , W ) ∈ S do
if S S S is a positive band then
res ← \gets ← RoundPositive( S , ρ P , W ) (S,\rho_P,\mathcal{W}) ( S , ρ P , W )
else
res ← \gets ← RoundNegative( S , ρ P , W ) (S,\rho_P,\mathcal{W}) ( S , ρ P , W )
if res = = = SUCCESS( I ) (I) ( I ) then
return I I I
res ← \gets ← IterativeGaussianLocalMaxima( G , { v x } x ∈ V , δ , T , R ) (G,{\{v_x\}}_{x\in V},\delta,T,R) ( G , { v x } x ∈ V , δ , T , R )
if res = = = SUCCESS( I ) (I) ( I ) then
return I I I
return FAIL
Band rounding. For each band S = S r , P S=S_{r,P} S = S r , P with witness W = W r , P \mathcal{W}=\mathcal{W}_{r,P} W = W r , P , Algorithm 3 calls Algorithm 4 if S S S is positive and Algorithm 5 if S S S is negative. Both procedures keep the vertices of S S S with large witness mass and round their vectors with VectorColoring.
If S S S is positive, then W = ( x ∅ , { x i } i ∈ S ) \mathcal{W}=(x_{\emptyset},\{x_i\}_{i\in S}) W = ( x ∅ , { x i } i ∈ S ) , and RoundPositive sets the threshold τ = 1 + 2 ρ P 3 − ε t h r \tau=\frac{1+2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ = 3 1 + 2 ρ P − ε thr and the set S ′ = { i ∈ S : ∥ x i ∥ 2 ≥ τ } S'=\{i\in S:\lVert x_i\rVert^2\ge\tau\} S ′ = { i ∈ S : ∥ x i ∥ 2 ≥ τ } . It splits S ′ S' S ′ into the set I 0 I_0 I 0 of isolated vertices of G [ S ′ ] G[S'] G [ S ′ ] and the rest S ′ ′ = S ′ ∖ I 0 S''=S'\setminus I_0 S ′′ = S ′ ∖ I 0 . The vectors z i = x i − ∥ x i ∥ x ∅ ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) z_i=\frac{x_i-\lVert x_i\rVert x_{\emptyset}}{\sqrt{\lVert x_i\rVert^2(1-\lVert x_i\rVert^2)}} z i = ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) x i − ∥ x i ∥ x ∅ for i ∈ S ′ ′ i\in S'' i ∈ S ′′ form an ordinary vector 1 / τ 1/\tau 1/ τ -coloring, as shown in Theorem 35 .
If S S S is negative, then W = ( x ∅ , { ( x i , 1 , x i , 2 ) } i ∈ S ) \mathcal{W}=(x_{\emptyset},\{(x_{i,1},x_{i,2})\}_{i\in S}) W = ( x ∅ , {( x i , 1 , x i , 2 ) } i ∈ S ) , and RoundNegative sets τ = 2 − 2 ρ P 3 − ε t h r \tau=\frac{2-2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ = 3 2 − 2 ρ P − ε thr and S ′ = { i ∈ S : ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 ≥ τ } S'=\{i\in S:\lVert x_{i,1}\rVert^2+\lVert x_{i,2}\rVert^2\ge\tau\} S ′ = { i ∈ S : ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 ≥ τ } . The vectors z i = x i , 1 − x i , 2 ∥ x i ∥ z_i=\frac{x_{i,1}-x_{i,2}}{\lVert x_i\rVert} z i = ∥ x i ∥ x i , 1 − x i , 2 for i ∈ S ′ i\in S' i ∈ S ′ form an ordinary vector 3 τ − 1 2 τ − 1 \frac{3\tau-1}{2\tau-1} 2 τ − 1 3 τ − 1 -coloring, as shown in Theorem 35 .
In both cases, the procedure returns SUCCESS with the resulting independent set if its size is at least T T T and FAIL otherwise.
Algorithm 4 RoundPositive
Require: positive band S S S , lower endpoint ρ P ≥ 0 \rho_P \ge0 ρ P ≥ 0 , witness W = ( x ∅ , { x i } i ∈ S ) \mathcal{W} = (x_{\emptyset}, \{x_i\}_{i\in S}) W = ( x ∅ , { x i } i ∈ S )
Ensure: SUCCESS(I I I ) or FAIL
τ ← 1 + 2 ρ P 3 − ε t h r \tau\leftarrow\frac{1+2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ ← 3 1 + 2 ρ P − ε thr
S ′ ← { i ∈ S : ∥ x i ∥ 2 ≥ τ } S' \leftarrow\{i \in S : \lVert x_i\rVert^2 \ge\tau\} S ′ ← { i ∈ S : ∥ x i ∥ 2 ≥ τ }
I 0 ← { i ∈ S ′ : N S ′ ( i ) = ∅ } I_0 \leftarrow\{i \in S' : N_{S'}(i) = \emptyset\} I 0 ← { i ∈ S ′ : N S ′ ( i ) = ∅ }
S ′ ′ ← S ′ ∖ I 0 S'' \leftarrow S' \setminus I_0 S ′′ ← S ′ ∖ I 0
z i ← x i − ∥ x i ∥ 2 x ∅ ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) z_i \leftarrow\frac{x_i-\lVert x_i\rVert^2x_{\emptyset}}{\sqrt{\lVert x_i\rVert^2(1-\lVert x_i\rVert^2)}} z i ← ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) x i − ∥ x i ∥ 2 x ∅ for every i ∈ S ′ ′ i \in S'' i ∈ S ′′
I ′ ← VectorColoring ( G [ S ′ ′ ] , { z i } i ∈ S ′ ′ , 1 τ ) I' \leftarrow\operatorname{VectorColoring}(G[S''], \{z_i\}_{i\in S''}, \frac{1}{\tau}) I ′ ← VectorColoring ( G [ S ′′ ] , { z i } i ∈ S ′′ , τ 1 )
I ← I 0 ∪ I ′ I \leftarrow I_0 \cup I' I ← I 0 ∪ I ′
return SUCCESS(I I I ) if ∣ I ∣ ≥ T \lvert I\rvert\ge T ∣ I ∣ ≥ T , and FAIL otherwise
Algorithm 5 RoundNegative
Require: negative band S S S , upper endpoint ρ P ≤ 0 \rho_P \le0 ρ P ≤ 0 , witness W = ( x ∅ , { ( x i , 1 , x i , 2 ) } i ∈ S ) \mathcal{W} = (x_{\emptyset}, \{(x_{i,1},x_{i,2})\}_{i\in S}) W = ( x ∅ , {( x i , 1 , x i , 2 ) } i ∈ S )
Ensure: SUCCESS(I I I ) or FAIL
τ ← 2 − 2 ρ P 3 − ε t h r \tau\leftarrow\frac{2-2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ ← 3 2 − 2 ρ P − ε thr
S ′ ← { i ∈ S : ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 ≥ τ } S' \leftarrow\{i \in S : \lVert x_{i,1}\rVert^2+\lVert x_{i,2}\rVert^2 \ge\tau\} S ′ ← { i ∈ S : ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 ≥ τ }
z i ← x i , 1 − x i , 2 ∥ x i ∥ z_i \leftarrow\frac{x_{i,1}-x_{i,2}}{\lVert x_i\rVert} z i ← ∥ x i ∥ x i , 1 − x i , 2 for every i ∈ S ′ i \in S' i ∈ S ′
I ← VectorColoring ( G [ S ′ ] , { z i } i ∈ S ′ , 3 τ − 1 2 τ − 1 ) I \leftarrow\operatorname{VectorColoring}(G[S'], \{z_i\}_{i\in S'}, \frac{3\tau-1}{2\tau-1}) I ← VectorColoring ( G [ S ′ ] , { z i } i ∈ S ′ , 2 τ − 1 3 τ − 1 )
return SUCCESS(I I I ) if ∣ I ∣ ≥ T \lvert I\rvert\ge T ∣ I ∣ ≥ T , and FAIL otherwise
For simplicity, in the analysis below, we treat the second and fourth constraints of Theorem 34 as exact, namely ⟨ v i , v j ⟩ = − 1 2 \langle v_i,v_j\rangle=-\frac{1}{2} ⟨ v i , v j ⟩ = − 2 1 for every i j ∈ E ij \in E ij ∈ E and Γ r x ( X ) ⪰ 0 \Gamma_{rx}(X) \succeq0 Γ r x ( X ) ⪰ 0 for every r x ∈ E rx \in E r x ∈ E . The actual implementation satisfies these constraints only up to an additive error of ε S D P \varepsilon_{\mathrm{SDP}} ε SDP , as stated in Theorem 34. We justify the use of the exact constraints by showing that these errors do not affect our analysis. Both issues are handled in Appendix § 9 .§ 2 .
Guarantees of Failed Band Rounding
In this subsection, we show that a band whose rounding finds no large independent set is small. The argument has three steps: BuildVector stores a witness of large mass for every band (Theorem 34), the witness yields a large subset of the band with an ordinary vector coloring (Theorem 35 ), and if VectorColoring finds no large independent set in this subset, then the band is small (Theorem 37). These bounds are used in Subsection 5.4 in order to obtain a sharper tail bound on the neighborhood maximum.
We use the parameters of Table 3 ; note that H 0 = D + 2 H_0=D+2 H 0 = D + 2 and δ = n − 1 / ( D + 2 ) \delta=n^{-1/(D+2)} δ = n − 1/ ( D + 2 ) . Throughout, every vertex of G G G has degree at most n β n^\beta n β .
Lemma 35 (Guarantee of band rounding). Fix a band S = S r , P S=S_{r,P} S = S r , P . Let 0 < ε t h r ≤ 1 / 30 0<\varepsilon_{\mathrm{thr}}\le1/30 0 < ε thr ≤ 1/30 . Then S ′ S' S ′ satisfies ∣ S ′ ∣ ≥ ε t h r 2 ∣ S ∣ \lvert S'\rvert\ge\frac{\varepsilon_{\mathrm{thr}}}{2}\lvert S\rvert ∣ S ′ ∣ ≥ 2 ε thr ∣ S ∣ .
(i) Suppose that S S S is positive and ρ P ≤ 1 4 \rho_P \le\frac{1}{4} ρ P ≤ 4 1 , and let β + ∈ ( 0 , 1 ] \beta_+ \in(0,1] β + ∈ ( 0 , 1 ] satisfy Δ ( G [ S ′ ′ ] ) ≤ ∣ S ′ ′ ∣ β + \Delta(G[S^{\prime\prime}]) \le|S^{\prime\prime}|^{\beta_+} Δ ( G [ S ′′ ]) ≤ ∣ S ′′ ∣ β + . Whenever the internal call to \textsc{VectorColoring} meets its guarantee, \textsc{RoundPositive} constructs an independent set I I I satisfying ∣ I ∣ ≥ ( ε t h r ∣ S ∣ ) 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) |I| \ge(\varepsilon_{\mathrm{thr}}|S|)^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)} ∣ I ∣ ≥ ( ε thr ∣ S ∣ ) 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) .
(ii) Suppose that S S S is positive and ρ P > 1 4 \rho_P > \frac{1}{4} ρ P > 4 1 . Then G [ S ] G[S] G [ S ] is edgeless, and hence we have I = I 0 = S ′ I=I_0=S' I = I 0 = S ′ .
(iii) Suppose that S S S is negative, and let β − ∈ ( 0 , 1 ] \beta_- \in(0,1] β − ∈ ( 0 , 1 ] satisfy Δ ( G [ S ′ ] ) ≤ ∣ S ′ ∣ β − \Delta(G[S']) \le|S'|^{\beta_-} Δ ( G [ S ′ ]) ≤ ∣ S ′ ∣ β − . Whenever the internal call to \textsc{VectorColoring} meets its guarantee, \textsc{RoundNegative} constructs an independent set I I I satisfying ∣ I ∣ ≥ ( ε t h r ∣ S ∣ ) 2 ( 1 + β − ) ( 2 τ − 1 ) ( 4 + 3 β − ) τ − ( 2 + β − ) − O ( h ) − o ( 1 ) |I| \ge(\varepsilon_{\mathrm{thr}}|S|)^{\frac{2(1+\beta_-)(2\tau-1)}{(4+3\beta_-)\tau-(2+\beta_-)}-O(h)-o(1)} ∣ I ∣ ≥ ( ε thr ∣ S ∣ ) ( 4 + 3 β − ) τ − ( 2 + β − ) 2 ( 1 + β − ) ( 2 τ − 1 ) − O ( h ) − o ( 1 ) .
Proof. First suppose that S S S is positive. From the definitions S ′ = { j ∈ S : ∥ x j ∥ 2 ≥ τ } S'=\{j\in S:\|x_j\|^2\ge\tau\} S ′ = { j ∈ S : ∥ x j ∥ 2 ≥ τ } and τ = 1 + 2 ρ P 3 − ε t h r \tau=\frac{1+2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ = 3 1 + 2 ρ P − ε thr , we have
∑ j ∈ S ∥ x j ∥ 2 ≤ ∣ S ′ ∣ + τ ∣ S ∖ S ′ ∣ ≤ ∣ S ′ ∣ + τ ∣ S ∣ . \sum_{j\in S}\|x_j\|^2 \le|S'|+\tau|S\setminus S'| \le|S'|+\tau|S|. j ∈ S ∑ ∥ x j ∥ 2 ≤ ∣ S ′ ∣ + τ ∣ S ∖ S ′ ∣ ≤ ∣ S ′ ∣ + τ ∣ S ∣.
Since the output of \textsc{BuildVector} satisfies
∑ j ∈ S ∥ x j ∥ 2 ≥ 1 + 2 ρ P 3 ∣ S ∣ − ε S D P = ( τ + ε t h r ) ∣ S ∣ − ε S D P , \sum_{j\in S}\|x_j\|^2 \ge\frac{1+2\rho_P}{3}|S|-\varepsilon_{\mathrm{SDP}}=(\tau+\varepsilon_{\mathrm{thr}})|S|-\varepsilon_{\mathrm{SDP}}, j ∈ S ∑ ∥ x j ∥ 2 ≥ 3 1 + 2 ρ P ∣ S ∣ − ε SDP = ( τ + ε thr ) ∣ S ∣ − ε SDP ,
we have ( τ + ε t h r ) ∣ S ∣ − ε S D P ≤ ∣ S ′ ∣ + τ ∣ S ∣ (\tau+\varepsilon_{\mathrm{thr}})|S|-\varepsilon_{\mathrm{SDP}}\le|S'|+\tau|S| ( τ + ε thr ) ∣ S ∣ − ε SDP ≤ ∣ S ′ ∣ + τ ∣ S ∣ , which gives ∣ S ′ ∣ ≥ ε t h r ∣ S ∣ − ε S D P ≥ ε t h r 2 ∣ S ∣ |S'|\ge\varepsilon_{\mathrm{thr}}|S|-\varepsilon_{\mathrm{SDP}}\ge\frac{\varepsilon_{\mathrm{thr}}}{2}|S| ∣ S ′ ∣ ≥ ε thr ∣ S ∣ − ε SDP ≥ 2 ε thr ∣ S ∣ for nonempty S S S , since ε S D P = e − 3 n ≤ ε t h r 2 \varepsilon_{\mathrm{SDP}}=e^{-3n}\le\frac{\varepsilon_{\mathrm{thr}}}{2} ε SDP = e − 3 n ≤ 2 ε thr for sufficiently large n n n . In the negative case, we can similarly show that ∣ S ′ ∣ ≥ ε t h r 2 ∣ S ∣ |S'|\ge\frac{\varepsilon_{\mathrm{thr}}}{2}|S| ∣ S ′ ∣ ≥ 2 ε thr ∣ S ∣ .
Suppose that S S S is positive and ρ P > 1 4 \rho_P>\frac{1}{4} ρ P > 4 1 . We have ⟨ v r , v j ⟩ + ⟨ v r , v j ′ ⟩ ≥ 2 ρ P > 1 2 \langle v_r,v_j\rangle+\langle v_r,v_{j'}\rangle\ge2\rho_P>\frac{1}{2} ⟨ v r , v j ⟩ + ⟨ v r , v j ′ ⟩ ≥ 2 ρ P > 2 1 for every edge j j ′ ∈ E [ S ] jj'\in E[S] j j ′ ∈ E [ S ] with r ∉ { j , j ′ } r\notin\{j,j'\} r ∈ / { j , j ′ } , contradicting the third constraint of Theorem 34. Moreover, for the fixed vertex r r r and every j ∈ N S ( r ) j\in N_S(r) j ∈ N S ( r ) , ⟨ v r , v j ⟩ ≤ − 1 2 + ε S D P < 1 4 \langle v_r,v_j\rangle\le-\frac{1}{2}+\varepsilon_{\mathrm{SDP}}<\frac{1}{4} ⟨ v r , v j ⟩ ≤ − 2 1 + ε SDP < 4 1 holds, which contradicts the assumption ρ P > 1 4 \rho_P>\frac{1}{4} ρ P > 4 1 . Thus, G [ S ] G[S] G [ S ] is edgeless and we can take an independent set I I I as I 0 = S ′ I_0=S' I 0 = S ′ .
Construction of a vector coloring for a positive band (ρ P ≤ 1 4 \rho_P\le\frac{1}{4} ρ P ≤ 4 1 ). Within this band, τ = 1 + 2 ρ P 3 − ε t h r ≤ 1 2 − ε t h r < 1 2 \tau=\frac{1+2\rho_P}{3}-\varepsilon_{\mathrm{thr}}\le\frac{1}{2}-\varepsilon_{\mathrm{thr}}<\frac{1}{2} τ = 3 1 + 2 ρ P − ε thr ≤ 2 1 − ε thr < 2 1 . From the constraints of Θ 1 ( S ) \Theta_1(S) Θ 1 ( S ) , we have ⟨ x i , x j ⟩ = 0 \langle x_i,x_j\rangle=0 ⟨ x i , x j ⟩ = 0 and ⟨ x ∅ − x i , x ∅ − x j ⟩ ≥ 0 \langle x_\varnothing-x_i,x_\varnothing-x_j\rangle\ge0 ⟨ x ∅ − x i , x ∅ − x j ⟩ ≥ 0 for i j ∈ E [ S ] ij\in E[S] ij ∈ E [ S ] . Hence ∥ x i ∥ 2 + ∥ x j ∥ 2 ≤ 1 \|x_i\|^2+\|x_j\|^2\le1 ∥ x i ∥ 2 + ∥ x j ∥ 2 ≤ 1 holds. Every vertex of S ′ ′ S^{\prime\prime} S ′′ has a neighbor in S ′ S' S ′ (and hence in S ′ ′ S^{\prime\prime} S ′′ ). From the definition of S ′ S' S ′ , we have τ ≤ ∥ x i ∥ 2 , ∥ x j ∥ 2 \tau\le\|x_i\|^2,\|x_j\|^2 τ ≤ ∥ x i ∥ 2 , ∥ x j ∥ 2 , which gives τ ≤ ∥ x i ∥ 2 ≤ 1 − τ \tau\le\|x_i\|^2\le1-\tau τ ≤ ∥ x i ∥ 2 ≤ 1 − τ for every i ∈ S ′ ′ i\in S^{\prime\prime} i ∈ S ′′ .
For i ∈ S ′ ′ i\in S^{\prime\prime} i ∈ S ′′ , set z i = x i − ∥ x i ∥ 2 x ∅ ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) z_i=\frac{x_i-\|x_i\|^2x_\varnothing}{\sqrt{\|x_i\|^2(1-\|x_i\|^2)}} z i = ∥ x i ∥ 2 ( 1 − ∥ x i ∥ 2 ) x i − ∥ x i ∥ 2 x ∅ , which is a unit vector. For i j ∈ E [ S ′ ′ ] ij\in E[S^{\prime\prime}] ij ∈ E [ S ′′ ] ,
⟨ z i , z j ⟩ = − ∥ x i ∥ 2 1 − ∥ x i ∥ 2 ∥ x j ∥ 2 1 − ∥ x j ∥ 2 ≤ − τ 1 − τ = − 1 1 τ − 1 . \begin{aligned}
\langle z_i,z_j\rangle=-\sqrt{\frac{\|x_i\|^2}{1-\|x_i\|^2}}\sqrt{\frac{\|x_j\|^2}{1-\|x_j\|^2}} \\
&\le-\frac{\tau}{1-\tau} \\
&=-\frac{1}{\frac{1}{\tau}-1}.
\end{aligned} ⟨ z i , z j ⟩ = − 1 − ∥ x i ∥ 2 ∥ x i ∥ 2 1 − ∥ x j ∥ 2 ∥ x j ∥ 2 ≤ − 1 − τ τ = − τ 1 − 1 1 .
Therefore, { z i } i ∈ S ′ ′ \{z_i\}_{i\in S^{\prime\prime}} { z i } i ∈ S ′′ is a vector 1 τ \frac{1}{\tau} τ 1 -coloring, and τ < 1 2 \tau<\frac{1}{2} τ < 2 1 gives 1 τ > 2 \frac{1}{\tau}>2 τ 1 > 2 .
Construction of a vector coloring for a negative band. Recall that x i = x i , 1 + x i , 2 x_i=x_{i,1}+x_{i,2} x i = x i , 1 + x i , 2 . From the constraints of Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) , we have ⟨ x i , 1 , x i , 2 ⟩ = 0 \langle x_{i,1},x_{i,2}\rangle=0 ⟨ x i , 1 , x i , 2 ⟩ = 0 and ⟨ x ∅ , x i , c ⟩ = ∥ x i , c ∥ 2 \langle x_{\varnothing},x_{i,c}\rangle=\lVert x_{i,c}\rVert^2 ⟨ x ∅ , x i , c ⟩ = ∥ x i , c ∥ 2 . Hence ∥ x i , 1 − x i , 2 ∥ 2 = ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 = ∥ x i ∥ 2 \lVert x_{i,1}-x_{i,2}\rVert^2=\lVert x_{i,1}\rVert^2+\lVert x_{i,2}\rVert^2=\lVert x_i\rVert^2 ∥ x i , 1 − x i , 2 ∥ 2 = ∥ x i , 1 ∥ 2 + ∥ x i , 2 ∥ 2 = ∥ x i ∥ 2 . From the definition of S ′ S' S ′ , we have ∥ x i ∥ 2 ≥ τ \lVert x_i\rVert^2\ge\tau ∥ x i ∥ 2 ≥ τ for every i ∈ S ′ i\in S' i ∈ S ′ . Moreover, for a negative band we have τ = 2 − 2 ρ P 3 − ε t h r ≥ 2 3 − ε t h r > 1 2 \tau=\frac{2-2\rho_P}{3}-\varepsilon_{\mathrm{thr}}\ge\frac{2}{3}-\varepsilon_{\mathrm{thr}}>\frac{1}{2} τ = 3 2 − 2 ρ P − ε thr ≥ 3 2 − ε thr > 2 1 .
For i ∈ S ′ i\in S' i ∈ S ′ , set z i = x i , 1 − x i , 2 ∥ x i ∥ z_i=\frac{x_{i,1}-x_{i,2}}{\lVert x_i\rVert} z i = ∥ x i ∥ x i , 1 − x i , 2 , which is a unit vector. For i j ∈ E [ S ′ ] ij\in E[S'] ij ∈ E [ S ′ ] , the constraints of Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) give ⟨ x i , 1 − x i , 2 , x j , 1 − x j , 2 ⟩ = − ⟨ x i , x j ⟩ \langle x_{i,1}-x_{i,2},x_{j,1}-x_{j,2}\rangle=-\langle x_i,x_j\rangle ⟨ x i , 1 − x i , 2 , x j , 1 − x j , 2 ⟩ = − ⟨ x i , x j ⟩ and 0 ≤ ⟨ x ∅ − x i , x ∅ − x j ⟩ = 1 − ∥ x i ∥ 2 − ∥ x j ∥ 2 + ⟨ x i , x j ⟩ 0\le\langle x_{\varnothing}-x_i,x_{\varnothing}-x_j\rangle=1-\lVert x_i\rVert^2-\lVert x_j\rVert^2+\langle x_i,x_j\rangle 0 ≤ ⟨ x ∅ − x i , x ∅ − x j ⟩ = 1 − ∥ x i ∥ 2 − ∥ x j ∥ 2 + ⟨ x i , x j ⟩ , and hence ⟨ x i , x j ⟩ ≥ ∥ x i ∥ 2 + ∥ x j ∥ 2 − 1 \langle x_i,x_j\rangle\ge\lVert x_i\rVert^2+\lVert x_j\rVert^2-1 ⟨ x i , x j ⟩ ≥ ∥ x i ∥ 2 + ∥ x j ∥ 2 − 1 . Therefore,
⟨ z i , z j ⟩ = − ⟨ x i , x j ⟩ ∥ x i ∥ ∥ x j ∥ ≤ 1 − ∥ x i ∥ 2 − ∥ x j ∥ 2 ∥ x i ∥ ∥ x j ∥ ≤ 1 − 2 ∥ x i ∥ ∥ x j ∥ ∥ x i ∥ ∥ x j ∥ ( ∵ AM–GM inequality ) ≤ 1 τ − 2 ( ∵ ∥ x i ∥ ∥ x j ∥ ≥ τ ) = − 1 3 τ − 1 2 τ − 1 − 1 . \begin{aligned}
\langle z_i,z_j\rangle
&=-\frac{\langle x_i,x_j\rangle}{\lVert x_i\rVert\lVert x_j\rVert}\\
&\le\frac{1-\lVert x_i\rVert^2-\lVert x_j\rVert^2}{\lVert x_i\rVert\lVert x_j\rVert}\\
&\le\frac{1-2\lVert x_i\rVert\lVert x_j\rVert}{\lVert x_i\rVert\lVert x_j\rVert}
&&(\because\text{ AM--GM inequality})\\
&\le\frac{1}{\tau}-2
&&(\because\lVert x_i\rVert\lVert x_j\rVert\ge\tau)\\
&=-\frac{1}{\frac{3\tau-1}{2\tau-1}-1}.
\end{aligned} ⟨ z i , z j ⟩ = − ∥ x i ∥ ∥ x j ∥ ⟨ x i , x j ⟩ ≤ ∥ x i ∥ ∥ x j ∥ 1 − ∥ x i ∥ 2 − ∥ x j ∥ 2 ≤ ∥ x i ∥ ∥ x j ∥ 1 − 2 ∥ x i ∥ ∥ x j ∥ ≤ τ 1 − 2 = − 2 τ − 1 3 τ − 1 − 1 1 . ( ∵ AM–GM inequality ) ( ∵ ∥ x i ∥ ∥ x j ∥ ≥ τ )
Therefore, { z i } i ∈ S ′ \{z_i\}_{i\in S'} { z i } i ∈ S ′ is a vector 3 τ − 1 2 τ − 1 \frac{3\tau-1}{2\tau-1} 2 τ − 1 3 τ − 1 -coloring.
Now, we apply VectorColoring to the constructed vectors { z i } i ∈ S ′ ′ \{z_i\}_{i\in S''} { z i } i ∈ S ′′ with k = 1 τ k=\frac{1}{\tau} k = τ 1 for a positive band and { z i } i ∈ S ′ \{z_i\}_{i\in S'} { z i } i ∈ S ′ with k = 3 τ − 1 2 τ − 1 k=\frac{3\tau-1}{2\tau-1} k = 2 τ − 1 3 τ − 1 for a negative band. For a positive band, 1 3 − ε t h r ≤ τ < 1 2 \frac{1}{3}-\varepsilon_{\mathrm{thr}}\le\tau<\frac{1}{2} 3 1 − ε thr ≤ τ < 2 1 gives 2 < 1 τ ≤ 3 2<\frac{1}{\tau}\le3 2 < τ 1 ≤ 3 , and for a negative band, 2 3 − ε t h r ≤ τ < 1 \frac{2}{3}-\varepsilon_{\mathrm{thr}}\le\tau<1 3 2 − ε thr ≤ τ < 1 gives 2 < 3 τ − 1 2 τ − 1 ≤ 7 2 2<\frac{3\tau-1}{2\tau-1}\le\frac{7}{2} 2 < 2 τ − 1 3 τ − 1 ≤ 2 7 , where we used ε t h r ≤ 1 30 \varepsilon_{\mathrm{thr}}\le\frac{1}{30} ε thr ≤ 30 1 . Hence, k ∈ [ 2 , 4 ] k\in[2,4] k ∈ [ 2 , 4 ] in both cases. Let G ′ G' G ′ be the graph to which it is applied. VectorColoring runs the algorithm of Theorem 2 with k h k_h k h and β h ( G ′ ) \beta_h(G') β h ( G ′ ) , and in both regimes of Theorem 2 it returns, with probability at least 3 4 \frac{3}{4} 4 3 , an independent set of size at least ∣ V ( G ′ ) ∣ 1 − ( k h − 2 ) β h ( G ′ ) 2 + k h β h ( G ′ ) − o ( 1 ) \lvert V(G')\rvert^{1-\frac{(k_h-2)\beta_h(G')}{2+k_h\beta_h(G')}-o(1)} ∣ V ( G ′ ) ∣ 1 − 2 + k h β h ( G ′ ) ( k h − 2 ) β h ( G ′ ) − o ( 1 ) . Since the pair ( k h , β h ( G ′ ) ) (k_h,\beta_h(G')) ( k h , β h ( G ′ )) lies in the finite set K h × D h \mathcal{K}_h\times\mathcal{D}_h K h × D h , the threshold for ∣ V ( G ′ ) ∣ \lvert V(G')\rvert ∣ V ( G ′ )∣ and the o ( 1 ) o(1) o ( 1 ) term are uniform over all bands. The function ( k , β ′ ) ↦ ( k − 2 ) β ′ 2 + k β ′ (k,\beta')\mapsto\frac{(k-2)\beta'}{2+k\beta'} ( k , β ′ ) ↦ 2 + k β ′ ( k − 2 ) β ′ is increasing in β ′ \beta' β ′ , and since its numerator is linear in each argument and its denominator is at least 2, changing k ∈ [ 2 , 4 ] k\in[2,4] k ∈ [ 2 , 4 ] and β ′ ∈ [ 0 , 2 ] \beta'\in[0,2] β ′ ∈ [ 0 , 2 ] by at most h h h changes it by O ( h ) O(h) O ( h ) . Since k ≤ k h ≤ k + h k\le k_h\le k+h k ≤ k h ≤ k + h and β h ( G ′ ) ≤ β + + h \beta_h(G')\le\beta_++h β h ( G ′ ) ≤ β + + h , we obtain
( k h − 2 ) β h ( G ′ ) 2 + k h β h ( G ′ ) ≤ ( k − 2 ) β + 2 + k β + + O ( h ) . \frac{(k_h-2)\beta_h(G')}{2+k_h\beta_h(G')}
\le
\frac{(k-2)\beta_+}{2+k\beta_+}+O(h). 2 + k h β h ( G ′ ) ( k h − 2 ) β h ( G ′ ) ≤ 2 + k β + ( k − 2 ) β + + O ( h ) .
For k = 1 τ k=\frac{1}{\tau} k = τ 1 and k = 3 τ − 1 2 τ − 1 k=\frac{3\tau-1}{2\tau-1} k = 2 τ − 1 3 τ − 1 , the quantity 1 − ( k − 2 ) β ′ 2 + k β ′ 1-\frac{(k-2)\beta'}{2+k\beta'} 1 − 2 + k β ′ ( k − 2 ) β ′ equals 2 τ ( 1 + β ′ ) 2 τ + β ′ \frac{2\tau(1+\beta')}{2\tau+\beta'} 2 τ + β ′ 2 τ ( 1 + β ′ ) and 2 ( 1 + β ′ ) ( 2 τ − 1 ) ( 4 + 3 β ′ ) τ − ( 2 + β ′ ) \frac{2(1+\beta')(2\tau-1)}{(4+3\beta')\tau-(2+\beta')} ( 4 + 3 β ′ ) τ − ( 2 + β ′ ) 2 ( 1 + β ′ ) ( 2 τ − 1 ) , respectively.
For a positive band, applying VectorColoring to G [ S ′ ′ ] G[S''] G [ S ′′ ] , where S ′ ′ = S ′ ∖ I 0 S''=S'\setminus I_0 S ′′ = S ′ ∖ I 0 , we obtain an independent set I ′ I' I ′ with ∣ I ′ ∣ ≥ ∣ S ′ ′ ∣ 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) \lvert I'\rvert\ge\lvert S''\rvert^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)} ∣ I ′ ∣ ≥ ∣ S ′′ ∣ 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) . We return I = I ′ ∪ I 0 I=I'\cup I_0 I = I ′ ∪ I 0 , and its size is bounded as follows:
∣ I ∣ ≥ ∣ I 0 ∣ + ∣ S ′ ′ ∣ 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) ≥ ∣ I 0 ∣ 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) + ∣ S ′ ′ ∣ 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) ≥ ∣ S ′ ∣ 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) ≥ ( ε t h r 2 ∣ S ∣ ) 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) ≥ ( ε t h r ∣ S ∣ ) 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) . \begin{aligned}
\lvert I\rvert
&\ge\lvert I_0\rvert+\lvert S''\rvert^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)}\\
&\ge\lvert I_0\rvert^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)}
+\lvert S''\rvert^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)}\\
&\ge\lvert S'\rvert^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)} \\
&\ge\left(\frac{\varepsilon_{\mathrm{thr}}}{2}|S|\right)^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)} \\
&\ge\left(\varepsilon_{\mathrm{thr}}|S|\right)^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)}.
\end{aligned} ∣ I ∣ ≥ ∣ I 0 ∣ + ∣ S ′′ ∣ 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) ≥ ∣ I 0 ∣ 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) + ∣ S ′′ ∣ 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) ≥ ∣ S ′ ∣ 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) ≥ ( 2 ε thr ∣ S ∣ ) 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) ≥ ( ε thr ∣ S ∣ ) 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) .
For a negative band, applying VectorColoring to G [ S ′ ] G[S'] G [ S ′ ] , we obtain an independent set I I I of size at least
( ε t h r ∣ S ∣ ) 2 ( 1 + β − ) ( 2 τ − 1 ) ( 4 + 3 β − ) τ − ( 2 + β − ) − O ( h ) − o ( 1 ) . \left(\varepsilon_{\mathrm{thr}}|S|\right)^{\frac{2(1+\beta_-)(2\tau-1)}{(4+3\beta_-)\tau-(2+\beta_-)}-O(h)-o(1)}. ( ε thr ∣ S ∣ ) ( 4 + 3 β − ) τ − ( 2 + β − ) 2 ( 1 + β − ) ( 2 τ − 1 ) − O ( h ) − o ( 1 ) .
This completes the proof.
We express the band-size bound on the scale 2 log ∣ S ∣ t 2 \frac{2\log|S|}{t^2} t 2 2 l o g ∣ S ∣ . The capacity C D ( ρ ) C_D(\rho) C D ( ρ ) defined below gives the resulting upper bound, with an additive slack, when band rounding fails under its stated guarantees.
Definition 36 (Capacity). For D > 0 D>0 D > 0 and ρ ∈ [ − 1 2 , 1 ] \rho\in[-\frac{1}{2},1] ρ ∈ [ − 2 1 , 1 ] , define
C D ( ρ ) = { min { D + 2 , 1 2 ( 1 + 1 + 6 D ( D + 1 ) 1 − 2 ρ 1 − 4 ρ ) } , − 1 2 ≤ ρ ≤ 0 , min { D + 2 , 1 2 ( 1 + 1 + 6 D ( D + 1 ) 1 + 2 ρ ) } , 0 ≤ ρ ≤ 1 4 , D + 1 , 1 4 < ρ ≤ 1. C_D(\rho)=
\begin{cases}
\min\left\{D+2,\frac{1}{2}\left(1+\sqrt{1+6D(D+1)\frac{1-2\rho}{1-4\rho}}\right)\right\}, & -\frac{1}{2}\le\rho\le0,\\
\min\left\{D+2,\frac{1}{2}\left(1+\sqrt{1+\frac{6D(D+1)}{1+2\rho}}\right)\right\}, & 0\le\rho\le\frac{1}{4},\\
D+1, & \frac{1}{4}<\rho\le1.
\end{cases} C D ( ρ ) = ⎩ ⎨ ⎧ min { D + 2 , 2 1 ( 1 + 1 + 6 D ( D + 1 ) 1 − 4 ρ 1 − 2 ρ ) } , min { D + 2 , 2 1 ( 1 + 1 + 1 + 2 ρ 6 D ( D + 1 ) ) } , D + 1 , − 2 1 ≤ ρ ≤ 0 , 0 ≤ ρ ≤ 4 1 , 4 1 < ρ ≤ 1.
Lemma 37 (Band bound after failed rounding). Fix a band S = S r , P S=S_{r,P} S = S r , P . Suppose that the corresponding band-rounding procedure returns FAIL and that, whenever it makes its internal call to VectorColoring, that call meets its guarantee. Then, for every ρ ∈ P \rho\in P ρ ∈ P ,
∣ S ∣ ≤ exp ( ( C D ( ρ ) + η ) t 2 2 ) . |S|\le\exp\left((C_D(\rho)+\eta)\frac{t^2}{2}\right). ∣ S ∣ ≤ exp ( ( C D ( ρ ) + η ) 2 t 2 ) .
Proof. Let s = 2 log ∣ S ∣ t 2 s=\frac{2\log|S|}{t^2} s = t 2 2 l o g ∣ S ∣ . When s < D + 1 + η / 2 s<D+1+\eta/2 s < D + 1 + η /2 , we have s < C D ( ρ ) + η / 2 s<C_D(\rho)+\eta/2 s < C D ( ρ ) + η /2 . Thus, we assume that s ≥ D + 1 + η / 2 s\ge D+1+\eta/2 s ≥ D + 1 + η /2 below.
First suppose that S S S is positive and ρ P > 1 4 \rho_P>\frac{1}{4} ρ P > 4 1 . By Theorem 35 , ∣ I ∣ = ∣ S ′ ∣ ≥ ε t h r 2 ∣ S ∣ |I|=|S'|\ge\frac{\varepsilon_{\mathrm{thr}}}{2}|S| ∣ I ∣ = ∣ S ′ ∣ ≥ 2 ε thr ∣ S ∣ . Since ∣ I ∣ < T |I|<T ∣ I ∣ < T , we have ∣ S ∣ < 2 T / ε t h r ≤ e ( D + 1 + o ( 1 ) ) t 2 / 2 |S|<2T/\varepsilon_{\mathrm{thr}}\le e^{(D+1+o(1))t^2/2} ∣ S ∣ < 2 T / ε thr ≤ e ( D + 1 + o ( 1 )) t 2 /2 , which gives s < D + 1 + o ( 1 ) = C D ( ρ ) + o ( 1 ) s<D+1+o(1)=C_D(\rho)+o(1) s < D + 1 + o ( 1 ) = C D ( ρ ) + o ( 1 ) .
Below we consider bands with ρ P ∈ [ − 1 2 , 1 4 ] \rho_P\in[-\frac{1}{2},\frac{1}{4}] ρ P ∈ [ − 2 1 , 4 1 ] . By Theorem 35 , ∣ S ′ ∣ ≥ ε t h r 2 ∣ S ∣ |S'|\ge\frac{\varepsilon_{\mathrm{thr}}}{2}|S| ∣ S ′ ∣ ≥ 2 ε thr ∣ S ∣ , so ∣ S ′ ∣ = e ( s + o ( 1 ) ) t 2 / 2 |S'|=e^{(s+o(1))t^2/2} ∣ S ′ ∣ = e ( s + o ( 1 )) t 2 /2 . For a positive band, since the independent set returned does not reach the target size, we have
∣ I 0 ∣ < T ≤ e ( D + 1 ) t 2 / 2 = ∣ S ∣ e ( D + 1 − s ) t 2 / 2 . |I_0|<T\le e^{(D+1)t^2/2}=|S|e^{(D+1-s)t^2/2}. ∣ I 0 ∣ < T ≤ e ( D + 1 ) t 2 /2 = ∣ S ∣ e ( D + 1 − s ) t 2 /2 .
Thus,
∣ S ′ ′ ∣ = ∣ S ′ ∣ − ∣ I 0 ∣ ≥ ε t h r 2 ∣ S ∣ − ∣ S ∣ e ( D + 1 − s ) t 2 / 2 = ( ε t h r 2 − o ( 1 ) ) ∣ S ∣ = e ( s + o ( 1 ) ) t 2 / 2 . |S''|=|S'|-|I_0|\ge\frac{\varepsilon_{\mathrm{thr}}}{2}|S|-|S|e^{(D+1-s)t^2/2}=\left(\frac{\varepsilon_{\mathrm{thr}}}{2}-o(1)\right)|S|=e^{(s+o(1))t^2/2}. ∣ S ′′ ∣ = ∣ S ′ ∣ − ∣ I 0 ∣ ≥ 2 ε thr ∣ S ∣ − ∣ S ∣ e ( D + 1 − s ) t 2 /2 = ( 2 ε thr − o ( 1 ) ) ∣ S ∣ = e ( s + o ( 1 )) t 2 /2 .
Since Δ ( G ) ≤ e D t 2 / 2 \Delta(G)\le e^{Dt^2/2} Δ ( G ) ≤ e D t 2 /2 and ∣ S ′ ′ ∣ = e ( s + o ( 1 ) ) t 2 / 2 |S''|=e^{(s+o(1))t^2/2} ∣ S ′′ ∣ = e ( s + o ( 1 )) t 2 /2 for a positive band and ∣ S ′ ∣ = e ( s + o ( 1 ) ) t 2 / 2 |S'|=e^{(s+o(1))t^2/2} ∣ S ′ ∣ = e ( s + o ( 1 )) t 2 /2 for a negative band, we have Δ ( G [ S ′ ′ ] ) ≤ e D t 2 / 2 = ∣ S ′ ′ ∣ D / s + o ( 1 ) \Delta(G[S''])\le e^{Dt^2/2}=|S''|^{D/s+o(1)} Δ ( G [ S ′′ ]) ≤ e D t 2 /2 = ∣ S ′′ ∣ D / s + o ( 1 ) and Δ ( G [ S ′ ] ) ≤ e D t 2 / 2 = ∣ S ′ ∣ D / s + o ( 1 ) \Delta(G[S'])\le e^{Dt^2/2}=|S'|^{D/s+o(1)} Δ ( G [ S ′ ]) ≤ e D t 2 /2 = ∣ S ′ ∣ D / s + o ( 1 ) , respectively. Hence we may take β ± = D s + O ( h ) + o ( 1 ) \beta_\pm=\frac{D}{s}+O(h)+o(1) β ± = s D + O ( h ) + o ( 1 ) in Theorem 35 , which lies in ( 0 , 1 ] (0,1] ( 0 , 1 ] since s ≥ D + 1 s\ge D+1 s ≥ D + 1 .
Positive case. For a fixed τ ≤ 1 2 \tau\le\frac{1}{2} τ ≤ 2 1 , 2 τ ( 1 + β ′ ) 2 τ + β ′ \frac{2\tau(1+\beta')}{2\tau+\beta'} 2 τ + β ′ 2 τ ( 1 + β ′ ) is decreasing in β ′ \beta' β ′ . By Theorem 35 ,
∣ I ∣ ≥ ( ε t h r ∣ S ∣ ) 2 τ ( 1 + β + ) 2 τ + β + − O ( h ) − o ( 1 ) ≥ exp ( ( 2 τ s ( s + D ) 2 τ s + D − O ( h ) − o ( 1 ) ) t 2 2 ) . |I|\ge(\varepsilon_{\mathrm{thr}}|S|)^{\frac{2\tau(1+\beta_+)}{2\tau+\beta_+}-O(h)-o(1)}
\ge\exp\left(\left(\frac{2\tau s(s+D)}{2\tau s+D}-O(h)-o(1)\right)\frac{t^2}{2}\right). ∣ I ∣ ≥ ( ε thr ∣ S ∣ ) 2 τ + β + 2 τ ( 1 + β + ) − O ( h ) − o ( 1 ) ≥ exp ( ( 2 τ s + D 2 τ s ( s + D ) − O ( h ) − o ( 1 ) ) 2 t 2 ) .
Failure of the band-rounding procedure gives
exp ( ( 2 τ s ( s + D ) 2 τ s + D − O ( h ) − o ( 1 ) ) t 2 2 ) ≤ ∣ I ∣ ≤ T ≤ e ( D + 1 ) t 2 / 2 . \exp\left(\left(\frac{2\tau s(s+D)}{2\tau s+D}-O(h)-o(1)\right)\frac{t^2}{2}\right)\le|I|\le T\le e^{(D+1)t^2/2}. exp ( ( 2 τ s + D 2 τ s ( s + D ) − O ( h ) − o ( 1 ) ) 2 t 2 ) ≤ ∣ I ∣ ≤ T ≤ e ( D + 1 ) t 2 /2 .
Hence, we have s ≤ 1 2 ( 1 + 1 + 2 D ( D + 1 ) τ ) + O ( h ) + o ( 1 ) s\le\frac{1}{2}\left(1+\sqrt{1+\frac{2D(D+1)}{\tau}}\right)+O(h)+o(1) s ≤ 2 1 ( 1 + 1 + τ 2 D ( D + 1 ) ) + O ( h ) + o ( 1 ) . Substituting τ = 1 + 2 ρ P 3 − ε t h r \tau=\frac{1+2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ = 3 1 + 2 ρ P − ε thr ,
s ≤ 1 2 ( 1 + 1 + 6 D ( D + 1 ) 1 + 2 ρ P ) + O ( h + ε t h r ) + o ( 1 ) . s\le\frac{1}{2}\left(1+\sqrt{1+\frac{6D(D+1)}{1+2\rho_P}}\right)+O(h+\varepsilon_{\mathrm{thr}})+o(1). s ≤ 2 1 ( 1 + 1 + 1 + 2 ρ P 6 D ( D + 1 ) ) + O ( h + ε thr ) + o ( 1 ) .
Negative case. For a fixed 1 2 < τ < 1 \frac{1}{2}<\tau<1 2 1 < τ < 1 , 2 ( 1 + β ′ ) ( 2 τ − 1 ) ( 4 + 3 β ′ ) τ − ( 2 + β ′ ) \frac{2(1+\beta')(2\tau-1)}{(4+3\beta')\tau-(2+\beta')} ( 4 + 3 β ′ ) τ − ( 2 + β ′ ) 2 ( 1 + β ′ ) ( 2 τ − 1 ) is decreasing in β ′ \beta' β ′ . By Theorem 35 with β − \beta_- β − as above, and since the band-rounding procedure fails, we obtain
exp ( ( 2 ( 2 τ − 1 ) s ( s + D ) ( 4 s + 3 D ) τ − ( 2 s + D ) − O ( h ) − o ( 1 ) ) t 2 2 ) ≤ ∣ I ∣ ≤ T ≤ e ( D + 1 ) t 2 / 2 . \exp\left(\left(\frac{2(2\tau-1)s(s+D)}{(4s+3D)\tau-(2s+D)}-O(h)-o(1)\right)\frac{t^2}{2}\right)\le|I|\le T\le e^{(D+1)t^2/2}. exp ( ( ( 4 s + 3 D ) τ − ( 2 s + D ) 2 ( 2 τ − 1 ) s ( s + D ) − O ( h ) − o ( 1 ) ) 2 t 2 ) ≤ ∣ I ∣ ≤ T ≤ e ( D + 1 ) t 2 /2 .
Thus, we have s ≤ 1 2 ( 1 + 1 + 2 D ( D + 1 ) ( 3 τ − 1 ) 2 τ − 1 ) + O ( h ) + o ( 1 ) s\le\frac{1}{2}\left(1+\sqrt{1+\frac{2D(D+1)(3\tau-1)}{2\tau-1}}\right)+O(h)+o(1) s ≤ 2 1 ( 1 + 1 + 2 τ − 1 2 D ( D + 1 ) ( 3 τ − 1 ) ) + O ( h ) + o ( 1 ) . Substituting τ = 2 − 2 ρ P 3 − ε t h r \tau=\frac{2-2\rho_P}{3}-\varepsilon_{\mathrm{thr}} τ = 3 2 − 2 ρ P − ε thr ,
s ≤ 1 2 ( 1 + 1 + 6 D ( D + 1 ) ( 1 − 2 ρ P ) 1 − 4 ρ P ) + O ( h + ε t h r ) + o ( 1 ) . s\le\frac{1}{2}\left(1+\sqrt{1+\frac{6D(D+1)(1-2\rho_P)}{1-4\rho_P}}\right)+O(h+\varepsilon_{\mathrm{thr}})+o(1). s ≤ 2 1 ( 1 + 1 + 1 − 4 ρ P 6 D ( D + 1 ) ( 1 − 2 ρ P ) ) + O ( h + ε thr ) + o ( 1 ) .
Moreover, in either case, ∣ S ∣ ≤ n |S|\le n ∣ S ∣ ≤ n gives s ≤ D + 2 s\le D+2 s ≤ D + 2 . Therefore, we obtain s ≤ C D ( ρ P ) + O ( h + ε t h r ) + o ( 1 ) ≤ C D ( ρ P ) + η / 2 s\le C_D(\rho_P)+O(h+\varepsilon_{\mathrm{thr}})+o(1)\le C_D(\rho_P)+\eta/2 s ≤ C D ( ρ P ) + O ( h + ε thr ) + o ( 1 ) ≤ C D ( ρ P ) + η /2 by the choice of ε t h r \varepsilon_{\mathrm{thr}} ε thr and h h h in Subsection 5.2 and for sufficiently large n n n . For any ρ ∈ P \rho\in P ρ ∈ P , we have ∣ ρ − ρ P ∣ = o ( 1 ) |\rho-\rho_P|=o(1) ∣ ρ − ρ P ∣ = o ( 1 ) , and hence ∣ C D ( ρ ) − C D ( ρ P ) ∣ = o ( 1 ) ≤ η / 2 |C_D(\rho)-C_D(\rho_P)|=o(1)\le\eta/2 ∣ C D ( ρ ) − C D ( ρ P ) ∣ = o ( 1 ) ≤ η /2 for sufficiently large n n n . Thus, we have
∣ S ∣ ≤ exp ( ( C D ( ρ ) + η ) t 2 2 ) . |S|\le\exp\left((C_D(\rho)+\eta)\frac{t^2}{2}\right). ∣ S ∣ ≤ exp ( ( C D ( ρ ) + η ) 2 t 2 ) .
Tail Bound on the Neighborhood Maximum
With D a D_a D a and H H H as in Table 3 , we have H = H 0 + 3 η = D a + 2 H=H_0+3\eta=D_a+2 H = H 0 + 3 η = D a + 2 .
In this subsection, we prove Theorem 38, which strengthens the tail bound on the neighborhood maximum used to construct the large-measure convex set K 1 K_1 K 1 . Starting from the threshold that follows from the degree bound alone, we lower the neighborhood threshold step by step along the nested cores. The induction simultaneously controls projection thresholds over the entire core (the global bound) and within the neighborhood of each vertex (the local bound). At each step, the previous local bound, applied at each candidate vertex, supplies the large-measure convex set for Ehrhard extrapolation, while the previous global bound ensures that the Gaussian projections of the possible transmission targets stay below the previous thresholds. After the preparation, the resulting tail bound is used in the improved lifting argument in Subsection 5.5.
For a vertex r ∈ V r\in V r ∈ V , we use ρ x \rho_x ρ x and u x u_x u x of Subsection 3.2.
Put ε t = exp ( − ( 1 4 + η ) t 2 2 ) \varepsilon_t=\exp\left(-\left(\frac{1}{4}+\eta\right)\frac{t^2}{2}\right) ε t = exp ( − ( 4 1 + η ) 2 t 2 ) . For i ≥ 0 i\ge0 i ≥ 0 , define the global and local events
G i = { ∃ x ∈ F i + 1 : ∣ ⟨ g , u x ⟩ ∣ > q A i ( ρ x ) t } , L i = { M F i + 1 ( r ) > A i t } . \mathcal{G}_i=\left\{\exists x\in F_{i+1}:|\langle g,u_x\rangle|>q_{A_i}(\rho_x)t\right\},\qquad\mathcal{L}_i=\left\{M_{F_{i+1}}(r)>A_i t\right\}. G i = { ∃ x ∈ F i + 1 : ∣ ⟨ g , u x ⟩ ∣ > q A i ( ρ x ) t } , L i = { M F i + 1 ( r ) > A i t } .
where the threshold function q A i ( ρ ) q_{A_i}(\rho) q A i ( ρ ) is defined in Subsubsection 5.4.1. Note that these events depend on r r r .
Lemma 38. Suppose, for every vertex r ∈ V r \in V r ∈ V , every P ∈ P P \in\mathcal{P} P ∈ P and every ρ ∈ P \rho\in P ρ ∈ P , ∣ S r , P ∣ ≤ exp ( ( C D ( ρ ) + η ) t 2 2 ) |S_{r,P}| \le\exp\left((C_D(\rho)+\eta)\frac{t^2}{2}\right) ∣ S r , P ∣ ≤ exp ( ( C D ( ρ ) + η ) 2 t 2 ) . Suppose that p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x) \le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 for every 1 ≤ j ≤ ℓ 1 \le j \le\ell 1 ≤ j ≤ ℓ and every x ∈ F j x \in F_j x ∈ F j . Then, for sufficiently large n n n , m 0 = o ( ℓ ) m_0=o(\ell) m 0 = o ( ℓ ) , and for every 0 ≤ i ≤ m 0 0 \le i \le m_0 0 ≤ i ≤ m 0 and every vertex r ∈ V r \in V r ∈ V , we have
P g [ G i ] ≤ ( i + 2 ) ε t and P g [ L i ] ≤ ( i + 2 ) ε t . \mathbb{P}_{g}[\mathcal{G}_i] \le(i+2)\varepsilon_t
\quad\text{and}\quad
\mathbb{P}_{g}[\mathcal{L}_i] \le(i+2)\varepsilon_t. P g [ G i ] ≤ ( i + 2 ) ε t and P g [ L i ] ≤ ( i + 2 ) ε t .
We first show how a tail bound on the neighborhood maximum yields a large-measure convex set for Ehrhard extrapolation. This conversion lemma is used both within the preparatory induction for the tail bound on the neighborhood maximum and in the improved lifting argument.
Lemma 39. Let F ⊆ V F \subseteq V F ⊆ V and x ∈ V x \in V x ∈ V with ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 . Let A ∈ R A \in\mathbb{R} A ∈ R and c ≥ 0 c \ge0 c ≥ 0 , and let
K 1 ( A , c ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( A − c ∣ ξ y ∣ ) t ∀ y ∈ N F ( x ) } . K_1(A,c)=\left\{g:\langle g_{\perp},w_y\rangle\le(A-c|\xi_y|)t\ \forall y\in N_F(x)\right\}. K 1 ( A , c ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( A − c ∣ ξ y ∣ ) t ∀ y ∈ N F ( x ) } . Then
1 − γ ( K 1 ( A , c ) ) ≤ P g [ M F ( x ) > A t ] P [ Z ∈ [ c t , c t + 1 t ] ] . 1-\gamma(K_1(A,c))\le
\frac{\mathbb{P}_{g}[M_F(x)>At]}
{\mathbb{P}[Z\in[ct,ct+\frac{1}{t}]]}. 1 − γ ( K 1 ( A , c )) ≤ P [ Z ∈ [ c t , c t + t 1 ]] P g [ M F ( x ) > A t ] . where Z ∼ N ( 0 , 1 ) Z\sim N(0,1) Z ∼ N ( 0 , 1 ) . In particular, if κ ≥ 0 \kappa\ge0 κ ≥ 0 , c 2 + κ 2 < 1 4 c^2+\kappa^2<\frac14 c 2 + κ 2 < 4 1 and P g [ M F ( x ) > A t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_{g}[M_F(x)>At]\le t^{-4}\overline{\Phi}(t/2) P g [ M F ( x ) > A t ] ≤ t − 4 Φ ( t /2 ) , then, for sufficiently large t t t ,
γ ( K 1 ( A , c ) ) ≥ Φ ( κ t ) . \gamma(K_1(A,c))\ge\Phi(\kappa t). γ ( K 1 ( A , c )) ≥ Φ ( κ t ) .
Proof of Theorem 39. Recall the decomposition of v x y v_{xy} v x y in Equation (4) for y ∈ N F ( x ) y\in N_F(x) y ∈ N F ( x ) , and that ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_{\perp},w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ . Set
Z = ⟨ g , v x r ⟩ . Z=\langle g,v_{xr}\rangle. Z = ⟨ g , v x r ⟩ .
Then ⟨ g , v x y ⟩ = ξ y Z + ⟨ g ⊥ , w y ⟩ \langle g,v_{xy}\rangle=\xi_y Z+\langle g_{\perp},w_y\rangle ⟨ g , v x y ⟩ = ξ y Z + ⟨ g ⊥ , w y ⟩ for every y ∈ N F ( x ) y\in N_F(x) y ∈ N F ( x ) , and Z Z Z is a standard Gaussian independent of g ⊥ g_{\perp} g ⊥ .
For each g ∉ K 1 ( A , c ) g\notin K_1(A,c) g ∈ / K 1 ( A , c ) , choose y ∈ N F ( x ) y\in N_F(x) y ∈ N F ( x ) with ⟨ g ⊥ , w y ⟩ > ( A − c ∣ ξ y ∣ ) t \langle g_{\perp},w_y\rangle>(A-c|\xi_y|)t ⟨ g ⊥ , w y ⟩ > ( A − c ∣ ξ y ∣ ) t using the fixed total order on V V V , and choose s y ∈ { − 1 , 1 } s_y\in\{-1,1\} s y ∈ { − 1 , 1 } so that s y ξ y = ∣ ξ y ∣ s_y\xi_y=|\xi_y| s y ξ y = ∣ ξ y ∣ . Both choices depend only on g ⊥ g_{\perp} g ⊥ . If s y Z ∈ [ c t , c t + 1 t ] s_yZ\in[ct,ct+\frac{1}{t}] s y Z ∈ [ c t , c t + t 1 ] , then
⟨ g , v x y ⟩ = ξ y Z + ⟨ g ⊥ , w y ⟩ > c ∣ ξ y ∣ t + ( A − c ∣ ξ y ∣ ) t ( ∵ ξ y Z = ∣ ξ y ∣ s y Z ≥ c ∣ ξ y ∣ t ) = A t . \begin{aligned}
\langle g,v_{xy}\rangle
&=\xi_y Z+\langle g_{\perp},w_y\rangle\\
&>c|\xi_y|t+(A-c|\xi_y|)t
&&\left(\because\ \xi_yZ=|\xi_y|s_yZ\ge c|\xi_y|t\right)\\
&=At.
\end{aligned} ⟨ g , v x y ⟩ = ξ y Z + ⟨ g ⊥ , w y ⟩ > c ∣ ξ y ∣ t + ( A − c ∣ ξ y ∣ ) t = A t . ( ∵ ξ y Z = ∣ ξ y ∣ s y Z ≥ c ∣ ξ y ∣ t )
Since y y y and s y s_y s y depend only on g ⊥ g_{\perp} g ⊥ , Z Z Z is independent of g ⊥ g_{\perp} g ⊥ , and s y Z s_yZ s y Z is a standard Gaussian for each fixed sign s y s_y s y , we obtain
( 1 − γ ( K 1 ( A , c ) ) ) P [ Z ∈ [ c t , c t + 1 t ] ] = P g [ g ∉ K 1 ( A , c ) and s y Z ∈ [ c t , c t + 1 t ] ] ≤ P g [ M F ( x ) > A t ] . \left(1-\gamma(K_1(A,c))\right)\mathbb{P}\left[Z\in\left[ct,ct+\frac{1}{t}\right]\right]
=\mathbb{P}_{g}\left[g\notin K_1(A,c)\text{ and }s_yZ\in\left[ct,ct+\frac{1}{t}\right]\right]
\le\mathbb{P}_{g}[M_F(x)>At]. ( 1 − γ ( K 1 ( A , c )) ) P [ Z ∈ [ c t , c t + t 1 ] ] = P g [ g ∈ / K 1 ( A , c ) and s y Z ∈ [ c t , c t + t 1 ] ] ≤ P g [ M F ( x ) > A t ] .
Dividing by P [ Z ∈ [ c t , c t + 1 t ] ] > 0 \mathbb{P}\left[Z\in\left[ct,ct+\frac{1}{t}\right]\right]>0 P [ Z ∈ [ c t , c t + t 1 ] ] > 0 proves the claim.
If moreover κ ≥ 0 \kappa\ge0 κ ≥ 0 , c 2 + κ 2 < 1 4 c^2+\kappa^2 < \frac{1}{4} c 2 + κ 2 < 4 1 and P g [ M F ( x ) > A t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_{g}[M_F(x)>At]\le t^{-4}\overline{\Phi}(t/2) P g [ M F ( x ) > A t ] ≤ t − 4 Φ ( t /2 ) , then 0 ≤ c < 1 2 0\le c<\frac{1}{2} 0 ≤ c < 2 1 and κ < 1 2 \kappa<\frac{1}{2} κ < 2 1 , and hence, for t ≥ 1 t\ge1 t ≥ 1 , Theorem 14 gives
P [ Z ∈ [ c t , c t + 1 t ] ] ≥ 1 t ϕ ( c t + 1 t ) = 1 2 π t e − c 2 t 2 / 2 − c − 1 / ( 2 t 2 ) ≥ e − 1 2 π t e − c 2 t 2 / 2 . \mathbb{P}[Z\in[ct,ct+\tfrac{1}{t}]]\ge\frac{1}{t}\phi(ct+\tfrac{1}{t})=\frac{1}{\sqrt{2\pi}t}e^{-c^2t^2/2-c-1/(2t^2)}\ge\frac{e^{-1}}{\sqrt{2\pi}t}e^{-c^2t^2/2}. P [ Z ∈ [ c t , c t + t 1 ]] ≥ t 1 ϕ ( c t + t 1 ) = 2 π t 1 e − c 2 t 2 /2 − c − 1/ ( 2 t 2 ) ≥ 2 π t e − 1 e − c 2 t 2 /2 .
Therefore, for sufficiently large t t t ,
1 − γ ( K 1 ( A , c ) ) ≤ t − 4 Φ ‾ ( t / 2 ) P [ Z ∈ [ c t , c t + 1 t ] ] ≤ e 2 π t − 3 e c 2 t 2 / 2 Φ ‾ ( t / 2 ) ≤ e 2 π t − 3 e ( 1 / 4 − κ 2 ) t 2 / 2 Φ ‾ ( t / 2 ) ( ∵ c 2 < 1 4 − κ 2 ) ≤ e 2 π t − 3 Φ ‾ ( κ t ) ( ∵ Theorem 13 , κ < 1 2 ) ≤ Φ ‾ ( κ t ) . \begin{aligned}
1-\gamma(K_1(A,c))&\le\frac{t^{-4}\overline{\Phi}(t/2)}{\mathbb{P}[Z\in[ct,ct+\tfrac{1}{t}]]}\\
&\le e\sqrt{2\pi}\,t^{-3}e^{c^2t^2/2}\overline{\Phi}(t/2)\\
&\le e\sqrt{2\pi}\,t^{-3}e^{(1/4-\kappa^2)t^2/2}\overline{\Phi}(t/2) \qquad(\because\ c^2<\tfrac{1}{4}-\kappa^2)\\
&\le e\sqrt{2\pi}\,t^{-3}\overline{\Phi}(\kappa t) \qquad(\because\ \text{Theorem 13},\ \kappa<\tfrac{1}{2})\\
&\le\overline{\Phi}(\kappa t).
\end{aligned} 1 − γ ( K 1 ( A , c )) ≤ P [ Z ∈ [ c t , c t + t 1 ]] t − 4 Φ ( t /2 ) ≤ e 2 π t − 3 e c 2 t 2 /2 Φ ( t /2 ) ≤ e 2 π t − 3 e ( 1/4 − κ 2 ) t 2 /2 Φ ( t /2 ) ( ∵ c 2 < 4 1 − κ 2 ) ≤ e 2 π t − 3 Φ ( κ t ) ( ∵ Theorem 13 , κ < 2 1 ) ≤ Φ ( κ t ) .
Thresholds and Parameters
The threshold q c q_c q c below is the one obtained directly from the band-size bounds and a Gaussian union bound (see Equation (7) ). The profile f A f_A f A is a stronger threshold that we realize inductively, and q A = min { q c , f A } q_A=\min\{q_c,f_A\} q A = min { q c , f A } combines the region where the direct estimate suffices with the region where the propagation analysis is needed. Property (P1) below controls the change of the threshold in one step. For 1.968 ≤ A ≤ A 0 1.968\le A\le A_0 1.968 ≤ A ≤ A 0 and ρ ∈ [ − 1 2 , 1 ] \rho\in[-\frac{1}{2},1] ρ ∈ [ − 2 1 , 1 ] , we define q c ( ρ ) = ( 1 − ρ 2 ) ( C D a ( ρ ) + 1 4 ) q_c(\rho)=\sqrt{(1-\rho^2)(C_{D_a}(\rho)+\frac{1}{4})} q c ( ρ ) = ( 1 − ρ 2 ) ( C D a ( ρ ) + 4 1 ) and q A ( ρ ) = min { q c ( ρ ) , f A ( ρ ) } q_A(\rho)=\min\{q_c(\rho),f_A(\rho)\} q A ( ρ ) = min { q c ( ρ ) , f A ( ρ )} , where
f A ( ρ ) = f ~ ( A ) 1 − ρ 2 − 0.714 ζ 1 + 0.826 ζ 1 2 − 0.8 ζ 2 + 0.45 ζ 2 2 + 0.39 ζ 3 + 0.65 ζ 3 2 , f_A(\rho)=\widetilde{f}(A)\sqrt{1-\rho^2}-0.714\zeta_1+0.826\zeta_1^2-0.8\zeta_2+0.45\zeta_2^2+0.39\zeta_3+0.65\zeta_3^2, f A ( ρ ) = f ( A ) 1 − ρ 2 − 0.714 ζ 1 + 0.826 ζ 1 2 − 0.8 ζ 2 + 0.45 ζ 2 2 + 0.39 ζ 3 + 0.65 ζ 3 2 ,
where f ~ ( A ) = 2.466 + 0.52 ( A − 2 ) + 2.46 ( A − 2 ) 2 \widetilde{f}(A)=2.466+0.52(A-2)+2.46(A-2)^2 f ( A ) = 2.466 + 0.52 ( A − 2 ) + 2.46 ( A − 2 ) 2 , ζ 1 = ( − 1 40 − ρ ) + \zeta_1=(-\frac{1}{40}-\rho)_+ ζ 1 = ( − 40 1 − ρ ) + , ζ 2 = ( ρ − 1 50 ) + \zeta_2=(\rho-\frac{1}{50})_+ ζ 2 = ( ρ − 50 1 ) + , ζ 3 = ( ρ − 1 4 ) + \zeta_3=(\rho-\frac{1}{4})_+ ζ 3 = ( ρ − 4 1 ) + and x + = max { x , 0 } x_+=\max\{x,0\} x + = max { x , 0 } . We use the following two properties of f A f_A f A and q A q_A q A .
(P1) For 1.968 ≤ A ≤ A ′ ≤ A 0 1.968\le A\le A'\le A_0 1.968 ≤ A ≤ A ′ ≤ A 0 and ρ ∈ [ − 1 2 , 1 ] \rho\in[-\frac{1}{2},1] ρ ∈ [ − 2 1 , 1 ] ,
0 ≤ q A ′ ( ρ ) − q A ( ρ ) ≤ 2 ( A ′ − A ) . 0\le q_{A'}(\rho)-q_A(\rho)\le2(A'-A). 0 ≤ q A ′ ( ρ ) − q A ( ρ ) ≤ 2 ( A ′ − A ) .
Indeed, ∂ A f A ( ρ ) = f ~ ′ ( A ) 1 − ρ 2 \partial_A f_A(\rho)=\widetilde{f}'(A)\sqrt{1-\rho^2} ∂ A f A ( ρ ) = f ′ ( A ) 1 − ρ 2 with f ~ ′ ( A ) = 0.52 + 4.92 ( A − 2 ) ∈ [ 0 , 2 ] \widetilde{f}'(A)=0.52+4.92(A-2)\in[0,2] f ′ ( A ) = 0.52 + 4.92 ( A − 2 ) ∈ [ 0 , 2 ] for 1.968 ≤ A ≤ A 0 1.968\le A\le A_0 1.968 ≤ A ≤ A 0 , and q c q_c q c does not depend on A A A .
(P2) For A = A 0 A=A_0 A = A 0 , q c ≤ f A q_c\le f_A q c ≤ f A , that is, q A = q c q_A=q_c q A = q c (proved in Appendix § 10 ).
Furthermore, we use the following lemma, verified by computation; its proof is given in Appendix § 10 . Cases 1 and 2 of this lemma are used for the global and the local update, respectively.
Lemma 40. Let 1.968 ≤ A ≤ A 0 1.968\le A\le A_0 1.968 ≤ A ≤ A 0 , and let ρ ∈ [ − 1 2 , 1 ) \rho\in[-\frac{1}{2},1) ρ ∈ [ − 2 1 , 1 ) satisfy the assumption on ( A , ρ ) (A,\rho) ( A , ρ ) in one of the two cases in the table below. Then there exist ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) , independent of ρ ′ \rho' ρ ′ , such that
(i) a , μ , c , κ ≥ 0 a,\mu,c,\kappa\ge0 a , μ , c , κ ≥ 0 , b ∈ R b \in\mathbb{R} b ∈ R and 0 < θ < 14 0 < \theta< 14 0 < θ < 14 ;
(ii) a 2 + b 2 + μ 2 ≤ 1 − 10 − 12 a^2+b^2+\mu^2 \le1-10^{-12} a 2 + b 2 + μ 2 ≤ 1 − 1 0 − 12 and c 2 + κ 2 ≤ 1 4 − 10 − 20 c^2+\kappa^2 \le\frac{1}{4}-10^{-20} c 2 + κ 2 ≤ 4 1 − 1 0 − 20 ;
(iii) L 2 + Λ ( μ + θ ( μ + κ ) ) 2 ≥ J + 10 − 18 L^2+\Lambda(\mu+\theta(\mu+\kappa))^2 \ge J+10^{-18} L 2 + Λ ( μ + θ ( μ + κ ) ) 2 ≥ J + 1 0 − 18 ;
(iv) for every ρ ′ \rho' ρ ′ satisfying − 1 2 ≤ ρ ′ ≤ 1 2 − ρ -\frac{1}{2} \le\rho' \le\frac{1}{2}-\rho − 2 1 ≤ ρ ′ ≤ 2 1 − ρ and 3 4 − ρ 2 − ( ρ ′ ) 2 − ρ ρ ′ ≥ 0 \frac{3}{4}-\rho^2-(\rho')^2-\rho\rho' \ge0 4 3 − ρ 2 − ( ρ ′ ) 2 − ρ ρ ′ ≥ 0 ,
L ( 1 / 2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ / 2 ) + θ c ∣ ρ ′ + ρ / 2 ∣ 1 − ρ 2 + 3 2 ( 1 + θ ) a − 3 2 θ A ≥ q A ( ρ ′ ) + 3 ⋅ 10 − 6 \frac{L(1/2+\rho\rho')-(1+\theta)b(\rho'+\rho/2)+\theta c\lvert\rho'+\rho/2\rvert}{\sqrt{1-\rho^2}}+\frac{3}{2}(1+\theta)a-\frac{\sqrt{3}}{2}\theta A \ge q_A(\rho')+3\cdot10^{-6} 1 − ρ 2 L ( 1/2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ /2 ) + θ c ∣ ρ ′ + ρ /2 ∣ + 2 3 ( 1 + θ ) a − 2 3 θ A ≥ q A ( ρ ′ ) + 3 ⋅ 1 0 − 6 .
Here ( L , J , Λ ) (L,J,\Lambda) ( L , J , Λ ) are as follows:
Table 1 (PDF p. 43).
Condition (ii) supplies the measure estimates needed for Ehrhard extrapolation, condition (iii) provides sufficient exponential decay for the probability bounds, and condition (iv) ensures that the transmitted Gaussian projection exceeds the required threshold. The same parameter choice must apply to every admissible neighboring correlation ρ ′ \rho' ρ ′ , since the convex sets in the Ehrhard argument use one parameter choice for all neighbors of x x x . The computational verification certifies only the existence of such parameters.
In the rest of this subsection, for x ∈ V x \in V x ∈ V with ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , we assume that ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) are parameters given by Theorem 40 with ρ = ρ x \rho=\rho_x ρ = ρ x , in one of the two cases and for some value A A A of the parameter there, and we let ( L , J , Λ ) (L,J,\Lambda) ( L , J , Λ ) be the corresponding entries of its table.
Let x ∈ V x \in V x ∈ V satisfy ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 . Then every y ∈ N V ( x ) y \in N_V(x) y ∈ N V ( x ) satisfies
− 1 2 ≤ ρ y ≤ 1 2 − ρ x , 3 4 − ρ x 2 − ρ y 2 − ρ x ρ y ≥ 0 , -\frac{1}{2}\le\rho_y\le\frac{1}{2}-\rho_x,\qquad\frac{3}{4}-\rho_x^2-\rho_y^2-\rho_x\rho_y\ge0, − 2 1 ≤ ρ y ≤ 2 1 − ρ x , 4 3 − ρ x 2 − ρ y 2 − ρ x ρ y ≥ 0 ,
that is, ρ ′ = ρ y \rho'=\rho_y ρ ′ = ρ y satisfies the condition on ρ ′ \rho' ρ ′ in Theorem 40 with ρ = ρ x \rho=\rho_x ρ = ρ x . Indeed, ρ y ≥ − 1 2 \rho_y\ge-\frac{1}{2} ρ y ≥ − 2 1 and ρ x + ρ y ≤ 1 2 \rho_x+\rho_y\le\frac{1}{2} ρ x + ρ y ≤ 2 1 by Theorem 34, where if y = r y=r y = r we have ρ x = − 1 2 \rho_x=-\frac{1}{2} ρ x = − 2 1 and ρ y = 1 \rho_y=1 ρ y = 1 directly, and 3 4 − ρ x 2 − ρ y 2 − ρ x ρ y ≥ 0 \frac{3}{4}-\rho_x^2-\rho_y^2-\rho_x\rho_y\ge0 4 3 − ρ x 2 − ρ y 2 − ρ x ρ y ≥ 0 since Equation (5) and ( ξ y ) 2 ≤ 1 (\xi_y)^2\le1 ( ξ y ) 2 ≤ 1 give ( ρ y + ρ x / 2 ) 2 ≤ 3 4 ( 1 − ρ x 2 ) (\rho_y+\rho_x/2)^2\le\frac{3}{4}(1-\rho_x^2) ( ρ y + ρ x /2 ) 2 ≤ 4 3 ( 1 − ρ x 2 ) .
Gaussian Bounds
We first prove Theorem 41 , and then Theorem 42 , whose two assertions are used in the induction steps for the global and the local bound in the proof of Theorem 38. The two assertions differ in how the vertex x x x is chosen. In the first assertion, x x x is fixed, so the parent Gaussian projection ⟨ g , u x ⟩ \langle g,u_x\rangle ⟨ g , u x ⟩ is independent of the transverse Gaussian coordinates, and Ehrhard extrapolation applies. In the second assertion, x x x is the maximizer of M F j ( r ) M_{F_j}(r) M F j ( r ) and depends on g g g ; we keep the event that x x x attains the maximum, and bound its correlation with the event that the transmission from x x x to its children fails, using the additional SDP constraint and the Gaussian Hölder inequality. For x ∈ V x\in V x ∈ V , we define U x = span { v x y : y ∈ N V ( x ) } \mathcal{U}_x=\operatorname{span}\{v_{xy}:y\in N_V(x)\} U x = span { v x y : y ∈ N V ( x )} . Since v x y ⊥ v x v_{xy}\perp v_x v x y ⊥ v x for every y ∈ N V ( x ) y\in N_V(x) y ∈ N V ( x ) , we have U x ⊆ v x ⊥ \mathcal{U}_x\subseteq v_x^\perp U x ⊆ v x ⊥ .
Lemma 41. Let X X X be the Gram matrix of { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V , and let r x ∈ E r_x \in E r x ∈ E . Suppose that Γ r x ( X ) ⪰ 0 \Gamma_{r_x}(X) \succeq0 Γ r x ( X ) ⪰ 0 . Then ∣ ⟨ u 1 , u 2 ⟩ ∣ ≤ 1 2 ∥ u 1 ∥ ∥ u 2 ∥ |\langle u_1,u_2\rangle| \le\frac{1}{2}\lVert u_1\rVert\lVert u_2\rVert ∣ ⟨ u 1 , u 2 ⟩ ∣ ≤ 2 1 ∥ u 1 ∥ ∥ u 2 ∥ for all u 1 ∈ U r u_1 \in\mathcal{U}_r u 1 ∈ U r and u 2 ∈ U x u_2 \in\mathcal{U}_x u 2 ∈ U x .
Proof. Write u 1 = ∑ y ∈ N V ( r ) ( α 1 ) y v r y u_1 = \sum_{y\in N_V(r)}(\alpha_1)_y v_{r_y} u 1 = ∑ y ∈ N V ( r ) ( α 1 ) y v r y and u 2 = ∑ z ∈ N V ( x ) ( α 2 ) z v x z u_2 = \sum_{z\in N_V(x)}(\alpha_2)_z v_{x_z} u 2 = ∑ z ∈ N V ( x ) ( α 2 ) z v x z . By the definition of Γ r x ( X ) \Gamma_{r_x}(X) Γ r x ( X ) , for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R ,
0 ≤ ( α 1 , λ α 2 ) T Γ r x ( X ) ( α 1 , λ α 2 ) = ∥ u 1 ∥ 2 + 4 λ ⟨ u 1 , u 2 ⟩ + λ 2 ∥ u 2 ∥ 2 . 0 \leq(\alpha_1,\lambda\alpha_2)^{\mathsf T}\Gamma_{r_x}(X)(\alpha_1,\lambda\alpha_2) = \lVert u_1\rVert^2 + 4\lambda\langle u_1,u_2\rangle+ \lambda^2\lVert u_2\rVert^2. 0 ≤ ( α 1 , λ α 2 ) T Γ r x ( X ) ( α 1 , λ α 2 ) = ∥ u 1 ∥ 2 + 4 λ ⟨ u 1 , u 2 ⟩ + λ 2 ∥ u 2 ∥ 2 . Hence the discriminant of this quadratic in λ \lambda λ is nonpositive, giving 4 ⟨ u 1 , u 2 ⟩ 2 ≤ ∥ u 1 ∥ 2 ∥ u 2 ∥ 2 4\langle u_1,u_2\rangle^2 \leq\lVert u_1\rVert^2\lVert u_2\rVert^2 4 ⟨ u 1 , u 2 ⟩ 2 ≤ ∥ u 1 ∥ 2 ∥ u 2 ∥ 2 , as claimed.
Lemma 42. Fix a vertex r ∈ V r \in V r ∈ V . Let 1 ≤ j ≤ ℓ 1 \leq j \leq\ell 1 ≤ j ≤ ℓ and 1.968 ≤ A ≤ A ′ ≤ A 0 1.968 \leq A \leq A' \leq A_0 1.968 ≤ A ≤ A ′ ≤ A 0 , with A ′ − A ≤ w A' - A \leq w A ′ − A ≤ w . Assume that p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x) \leq e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 and P g [ M F j − 1 ( x ) > A ′ t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_g[M_{F_{j-1}}(x) > A't] \leq t^{-4}\overline{\Phi}(t/2) P g [ M F j − 1 ( x ) > A ′ t ] ≤ t − 4 Φ ( t /2 ) for every x ∈ F j x \in F_j x ∈ F j . In each of the following assertions, take the parameters ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) from the corresponding case of Theorem 40. Then, for sufficiently large t t t , the following hold.
(i) Let x ∈ F j x \in F_j x ∈ F j satisfy ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 and q A ( ρ x ) < q c ( ρ x ) q_A(\rho_x) < q_c(\rho_x) q A ( ρ x ) < q c ( ρ x ) . Let L L L be as in Case 1 of Theorem 40 with ρ = ρ x \rho= \rho_x ρ = ρ x . Then
P g [ ∣ ⟨ g , u x ⟩ ∣ > L ∥ u x ∥ t , ∣ ⟨ g , u y ⟩ ∣ ≤ q A ′ ( ρ y ) t ∀ y ∈ N F j − 1 ( x ) ] ≤ 2 Φ ‾ ( L t ) Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) . \mathbb{P}_g[|\langle g,u_x\rangle| > L\lVert u_x\rVert t, |\langle g,u_y\rangle| \leq q_{A'}(\rho_y)t\ \forall y \in N_{F_{j-1}}(x)] \leq2\overline{\Phi}(Lt)\overline{\Phi}((\mu+\theta(\mu+\kappa))t). P g [ ∣ ⟨ g , u x ⟩ ∣ > L ∥ u x ∥ t , ∣ ⟨ g , u y ⟩ ∣ ≤ q A ′ ( ρ y ) t ∀ y ∈ N F j − 1 ( x )] ≤ 2 Φ ( L t ) Φ (( μ + θ ( μ + κ )) t ) . (ii) Let X X X be the Gram matrix of { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V , and suppose that Γ r x ( X ) ⪰ 0 \Gamma_{r_x}(X) \succeq0 Γ r x ( X ) ⪰ 0 for every x ∈ N F j ( r ) x \in N_{F_j}(r) x ∈ N F j ( r ) , and take the parameters from Case 2 of Theorem 40. Put L = { M F j ( r ) > A t } \mathcal{L} = \{M_{F_j}(r) > At\} L = { M F j ( r ) > A t } . On L \mathcal{L} L , let x x x denote the maximizer of M F j ( r ) M_{F_j}(r) M F j ( r ) . Then, for all p 1 , p 2 > 1 p_1,p_2 > 1 p 1 , p 2 > 1 with ( p 1 − 1 ) ( p 2 − 1 ) ≥ 1 4 (p_1-1)(p_2-1) \geq\frac{1}{4} ( p 1 − 1 ) ( p 2 − 1 ) ≥ 4 1 ,
P g [ L , − ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t ∀ y ∈ N F j − 1 ( x ) ] ≤ ( ∣ N F j ( r ) ∣ Φ ‾ ( A t ) ) 1 − 1 / p 1 Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 P g [ L ] 1 / p 1 . \mathbb{P}_g[\mathcal{L},-\langle g,u_y\rangle\leq q_{A'}(\rho_y)t\ \forall y \in N_{F_{j-1}}(x)] \leq(|N_{F_j}(r)|\overline{\Phi}(At))^{1-1/p_1}\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2}\mathbb{P}_g[\mathcal{L}]^{1/p_1}. P g [ L , − ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t ∀ y ∈ N F j − 1 ( x )] ≤ ( ∣ N F j ( r ) ∣ Φ ( A t ) ) 1 − 1/ p 1 Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 P g [ L ] 1/ p 1 .
Proof.
First assertion. Let x ∈ F j x \in F_j x ∈ F j satisfy ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 . By the second equality of Equation (6) , every y ∈ N F j − 1 ( x ) y \in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) satisfies
u y = − 1 / 2 + ρ x ρ y 1 − ρ x 2 u x + 3 2 w y . u_y = -\frac{1/2+\rho_x\rho_y}{1-\rho_x^2}u_x + \frac{\sqrt{3}}{2}w_y. u y = − 1 − ρ x 2 1/2 + ρ x ρ y u x + 2 3 w y .
where u x ∈ span ( v r , v x ) u_x \in\operatorname{span}(v_r,v_x) u x ∈ span ( v r , v x ) and w y ∈ span ( v r , v x ) ⊥ w_y \in\operatorname{span}(v_r,v_x)^\perp w y ∈ span ( v r , v x ) ⊥ . Taking the inner product with g g g and using ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle= \langle g_\perp,w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ , we obtain
⟨ g , u y ⟩ = − 1 / 2 + ρ x ρ y 1 − ρ x 2 ⟨ g , u x ⟩ + 3 2 ⟨ g ⊥ , w y ⟩ . \langle g,u_y\rangle= -\frac{1/2+\rho_x\rho_y}{1-\rho_x^2}\langle g,u_x\rangle+ \frac{\sqrt{3}}{2}\langle g_\perp,w_y\rangle. ⟨ g , u y ⟩ = − 1 − ρ x 2 1/2 + ρ x ρ y ⟨ g , u x ⟩ + 2 3 ⟨ g ⊥ , w y ⟩ .
where ⟨ g , u x ⟩ \langle g,u_x\rangle ⟨ g , u x ⟩ is independent of g ⊥ g_\perp g ⊥ . Take ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) and L L L from either case of Theorem 40 with ρ = ρ x \rho= \rho_x ρ = ρ x , and put
B x = { ∀ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ ≤ ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( A ′ − c ∣ ξ y ∣ ) t − 1 + θ t } . \mathcal{B}_x = \left\{\forall y \in N_{F_{j-1}}(x): \langle g_\perp,w_y\rangle\leq(1+\theta)(\sqrt{3}a-b\xi_y)t-\theta(A'-c|\xi_y|)t-\frac{1+\theta}{t}\right\}. B x = { ∀ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ ≤ ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( A ′ − c ∣ ξ y ∣ ) t − t 1 + θ } .
Since B x \mathcal{B}_x B x is defined through ⟨ g ⊥ , w y ⟩ \langle g_\perp,w_y\rangle ⟨ g ⊥ , w y ⟩ only, it depends only on g ⊥ g_\perp g ⊥ . Moreover, if r x ∈ E r_x \in E r x ∈ E , then v x r ∈ U x v_{xr} \in\mathcal{U}_x v x r ∈ U x , and hence w y = v x y − ξ y v x r ∈ U x w_y = v_{xy}-\xi_yv_{xr} \in\mathcal{U}_x w y = v x y − ξ y v x r ∈ U x for every y ∈ N F j − 1 ( x ) y \in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) by Equation (4) . Since also w y ⊥ span ( v r , v x ) w_y \perp\operatorname{span}(v_r,v_x) w y ⊥ span ( v r , v x ) and ⟨ g ⊥ , w y ⟩ = ⟨ g , w y ⟩ \langle g_\perp,w_y\rangle=\langle g,w_y\rangle ⟨ g ⊥ , w y ⟩ = ⟨ g , w y ⟩ , in this case B x \mathcal{B}_x B x depends only on the Gaussian projection onto U x ∩ span ( v r , v x ) ⊥ \mathcal{U}_x\cap\operatorname{span}(v_r,v_x)^\perp U x ∩ span ( v r , v x ) ⊥ . We first bound P g [ B x ] \mathbb{P}_g[\mathcal{B}_x] P g [ B x ] . Since c 2 + κ 2 < 1 4 c^2+\kappa^2<\frac{1}{4} c 2 + κ 2 < 4 1 by (ii) of Theorem 40, and P g [ M F j − 1 ( x ) > A ′ t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_g[M_{F_{j-1}}(x)>A't]\le t^{-4}\overline{\Phi}(t/2) P g [ M F j − 1 ( x ) > A ′ t ] ≤ t − 4 Φ ( t /2 ) , Theorem 39 with F = F j − 1 F=F_{j-1} F = F j − 1 and A ′ A' A ′ in place of A A A gives, for sufficiently large t t t ,
P g [ ∀ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ ≤ ( A ′ − c ∣ ξ y ∣ ) t ] ≥ Φ ( κ t ) . \mathbb{P}_g\left[\forall y\in N_{F_{j-1}}(x):\langle g_\perp,w_y\rangle\le(A'-c|\xi_y|)t\right]\ge\Phi(\kappa t). P g [ ∀ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ ≤ ( A ′ − c ∣ ξ y ∣ ) t ] ≥ Φ ( κ t ) .
Moreover, since a ≥ 0 a\ge0 a ≥ 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 , we have, for sufficiently large t t t ,
γ ( K 0 ( a , b ) ) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ( ∵ citet‘Theorem 30‘ ) ≤ 2 π e 3 t 2 e − ( μ 2 + 10 − 12 ) t 2 / 2 ( ∵ (ii) of citet‘Theorem 40‘ ) ≤ Φ ‾ ( μ t ) . ( ∵ citet‘Theorem 12‘ ) \begin{aligned}
\gamma(K_0(a,b))&\le2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}} &&(\because\ \text{{citet}`Theorem 30`})\\
&\le2\pi e^3t^2e^{-(\mu^2+10^{-12})t^2/2} &&(\because\ \text{(ii) of {citet}`Theorem 40`})\\
&\le\overline{\Phi}(\mu t). &&(\because\ \text{{citet}`Theorem 12`})
\end{aligned} γ ( K 0 ( a , b )) ≤ 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ≤ 2 π e 3 t 2 e − ( μ 2 + 1 0 − 12 ) t 2 /2 ≤ Φ ( μ t ) . ( ∵ citet ‘Theorem 30‘ ) ( ∵ (ii) of citet ‘Theorem 40‘ ) ( ∵ citet ‘Theorem 12‘ )
Since ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_\perp,w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ , the sets K 0 K_0 K 0 and K 1 K_1 K 1 in Theorem 29 for Y = N F j − 1 ( x ) Y=N_{F_{j-1}}(x) Y = N F j − 1 ( x ) , z y = w y z_y=w_y z y = w y , A y = ( 3 a − b ξ y ) t − 1 t A_y=(\sqrt{3}a-b\xi_y)t-\frac{1}{t} A y = ( 3 a − b ξ y ) t − t 1 and B y = ( A ′ − c ∣ ξ y ∣ ) t B_y=(A'-c|\xi_y|)t B y = ( A ′ − c ∣ ξ y ∣ ) t are K 0 ( a , b ) K_0(a,b) K 0 ( a , b ) and K 1 ( A ′ , c ) K_1(A',c) K 1 ( A ′ , c ) . Hence Theorem 29 with μ 0 = μ t \mu_0=\mu t μ 0 = μ t and κ 0 = κ t \kappa_0=\kappa t κ 0 = κ t gives γ ( K − θ ) ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) \gamma(K_{-\theta})\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t) γ ( K − θ ) ≤ Φ (( μ + θ ( μ + κ )) t ) . Since K − θ = B x K_{-\theta}=\mathcal{B}_x K − θ = B x by definition, this means P g [ B x ] ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) \mathbb{P}_g[\mathcal{B}_x]\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t) P g [ B x ] ≤ Φ (( μ + θ ( μ + κ )) t ) for sufficiently large t t t , that is,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( A ′ − c ∣ ξ y ∣ ) t − 1 + θ t ] ≥ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) . \mathbb{P}_g\left[\exists y\in N_{F_{j-1}}(x):\langle g_\perp,w_y\rangle>(1+\theta)(\sqrt{3}a-b\xi_y)t-\theta(A'-c|\xi_y|)t-\frac{1+\theta}{t}\right]\ge\overline{\Phi}((\mu+\theta(\mu+\kappa))t). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( A ′ − c ∣ ξ y ∣ ) t − t 1 + θ ] ≥ Φ (( μ + θ ( μ + κ )) t ) .
Next, we show that B x \mathcal{B}_x B x occurs whenever − ⟨ g , u x ⟩ > L ∥ u x ∥ t -\langle g,u_x\rangle>L\lVert u_x\rVert t − ⟨ g , u x ⟩ > L ∥ u x ∥ t and ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t \langle g,u_y\rangle\le q_{A'}(\rho_y)t ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t for every y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) . Suppose that − ⟨ g , u x ⟩ > L ∥ u x ∥ t -\langle g,u_x\rangle>L\lVert u_x\rVert t − ⟨ g , u x ⟩ > L ∥ u x ∥ t . Then, for every y y y in the event above, we have
⟨ g , u y ⟩ t ≥ L ( 1 / 2 + ρ x ρ y ) 1 − ρ x 2 + 3 2 ⟨ g ⊥ , w y ⟩ t > L ( 1 / 2 + ρ x ρ y ) 1 − ρ x 2 + 3 2 ( ( 1 + θ ) ( 3 a − b ξ y ) − θ ( A ′ − c ∣ ξ y ∣ ) ) − 3 ( 1 + θ ) 2 t 2 = L ( 1 / 2 + ρ x ρ y ) − ( 1 + θ ) b ( ρ y + ρ x / 2 ) + θ c ∣ ρ y + ρ x / 2 ∣ 1 − ρ x 2 + 3 2 ( 1 + θ ) a − 3 2 θ A ′ − 3 ( 1 + θ ) 2 t 2 ≥ q A ( ρ y ) + 3 ⋅ 10 − 6 − 3 2 θ ( A ′ − A ) − 3 ( 1 + θ ) 2 t 2 ≥ q A ′ ( ρ y ) + 3 ⋅ 10 − 6 − ( 2 + 7 3 ) w − 15 3 2 t 2 > q A ′ ( ρ y ) , \begin{aligned}
\frac{\langle g,u_y\rangle}{t}
&\ge\frac{L(1/2+\rho_x\rho_y)}{\sqrt{1-\rho_x^2}}+\frac{\sqrt{3}}{2}\frac{\langle g_\perp,w_y\rangle}{t}\\
&> \frac{L(1/2+\rho_x\rho_y)}{\sqrt{1-\rho_x^2}}+\frac{\sqrt{3}}{2}\left((1+\theta)(\sqrt{3}a-b\xi_y)-\theta(A'-c|\xi_y|)\right)-\frac{\sqrt{3}(1+\theta)}{2t^2}\\
&=\frac{L(1/2+\rho_x\rho_y)-(1+\theta)b(\rho_y+\rho_x/2)+\theta c|\rho_y+\rho_x/2|}{\sqrt{1-\rho_x^2}}+\frac{3}{2}(1+\theta)a-\frac{\sqrt{3}}{2}\theta A'-\frac{\sqrt{3}(1+\theta)}{2t^2}\\
&\ge q_A(\rho_y)+3\cdot10^{-6}-\frac{\sqrt{3}}{2}\theta(A'-A)-\frac{\sqrt{3}(1+\theta)}{2t^2}\\
&\ge q_{A'}(\rho_y)+3\cdot10^{-6}-(2+7\sqrt{3})w-\frac{15\sqrt{3}}{2t^2}\\
&>q_{A'}(\rho_y),
\end{aligned} t ⟨ g , u y ⟩ ≥ 1 − ρ x 2 L ( 1/2 + ρ x ρ y ) + 2 3 t ⟨ g ⊥ , w y ⟩ > 1 − ρ x 2 L ( 1/2 + ρ x ρ y ) + 2 3 ( ( 1 + θ ) ( 3 a − b ξ y ) − θ ( A ′ − c ∣ ξ y ∣ ) ) − 2 t 2 3 ( 1 + θ ) = 1 − ρ x 2 L ( 1/2 + ρ x ρ y ) − ( 1 + θ ) b ( ρ y + ρ x /2 ) + θ c ∣ ρ y + ρ x /2∣ + 2 3 ( 1 + θ ) a − 2 3 θ A ′ − 2 t 2 3 ( 1 + θ ) ≥ q A ( ρ y ) + 3 ⋅ 1 0 − 6 − 2 3 θ ( A ′ − A ) − 2 t 2 3 ( 1 + θ ) ≥ q A ′ ( ρ y ) + 3 ⋅ 1 0 − 6 − ( 2 + 7 3 ) w − 2 t 2 15 3 > q A ′ ( ρ y ) ,
where the first inequality uses 1 2 + ρ x ρ y ≥ 0 \frac{1}{2}+\rho_x\rho_y\ge0 2 1 + ρ x ρ y ≥ 0 , which holds since ρ x , ρ y ∈ [ − 1 2 , 1 ] \rho_x,\rho_y\in[-\frac{1}{2},1] ρ x , ρ y ∈ [ − 2 1 , 1 ] , the second holds since y y y is in the event above, the equality follows from Equation (5) , the third inequality follows from (iv) of Theorem 40, which applies since ρ ′ = ρ y \rho'=\rho_y ρ ′ = ρ y satisfies the condition on ρ ′ \rho' ρ ′ there by the bounds on ρ y \rho_y ρ y above, the fourth uses (P1), A ′ − A ≤ w A'-A\le w A ′ − A ≤ w and θ < 14 \theta<14 θ < 14 , and the last inequality holds for sufficiently large t t t since w = log t / t 2 w=\log t/t^2 w = log t / t 2 .
Hence − ⟨ g , u x ⟩ > L ∥ u x ∥ t -\langle g,u_x\rangle>L\lVert u_x\rVert t − ⟨ g , u x ⟩ > L ∥ u x ∥ t and ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t \langle g,u_y\rangle\le q_{A'}(\rho_y)t ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t for every y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) imply that the event above fails, that is, B x \mathcal{B}_x B x occurs.
Since B x \mathcal{B}_x B x depends only on g ⊥ g_{\perp} g ⊥ , which is independent of ⟨ g , u x ⟩ \langle g,u_x\rangle ⟨ g , u x ⟩ , we obtain
P g [ − ⟨ g , u x ⟩ > L ∥ u x ∥ t , ∀ y ∈ N F j − 1 ( x ) : ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t ] ≤ P g [ − ⟨ g , u x ⟩ > L ∥ u x ∥ t ] P g [ B x ] ≤ Φ ‾ ( L t ) Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) , \mathbb{P}_g[-\langle g,u_x\rangle>L\lVert u_x\rVert t,\ \forall y\in N_{F_{j-1}}(x):\langle g,u_y\rangle\le q_{A'}(\rho_y)t]\le\mathbb{P}_g[-\langle g,u_x\rangle>L\lVert u_x\rVert t]\mathbb{P}_g[\mathcal{B}_x]
\le\overline{\Phi}(Lt)\overline{\Phi}((\mu+\theta(\mu+\kappa))t), P g [ − ⟨ g , u x ⟩ > L ∥ u x ∥ t , ∀ y ∈ N F j − 1 ( x ) : ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t ] ≤ P g [ − ⟨ g , u x ⟩ > L ∥ u x ∥ t ] P g [ B x ] ≤ Φ ( L t ) Φ (( μ + θ ( μ + κ )) t ) ,
where the second inequality follows from the bound on P g [ B x ] \mathbb{P}_g[\mathcal{B}_x] P g [ B x ] above. Applying this also to − g -g − g gives the first assertion.
Second assertion. Let x ∈ N F j ( r ) x\in N_{F_j}(r) x ∈ N F j ( r ) . Then r x ∈ E rx\in E r x ∈ E and ρ x = − 1 2 \rho_x=-\frac{1}{2} ρ x = − 2 1 , so ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , and ⟨ g , v r x ⟩ > A t \langle g,v_{rx}\rangle>At ⟨ g , v r x ⟩ > A t means − ⟨ − g , u x ⟩ > A ∥ u x ∥ t -\langle-g,u_x\rangle>A\lVert u_x\rVert t − ⟨ − g , u x ⟩ > A ∥ u x ∥ t , where L = A L=A L = A in Case 2. Let B x − \mathcal{B}_x^{-} B x − be the event B x \mathcal{B}_x B x above for the parameters of Case 2, with − g -g − g in place of g g g . By the bound on P g [ B x ] \mathbb{P}_g[\mathcal{B}_x] P g [ B x ] above and the symmetry of g g g , we have P g [ B x − ] ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) \mathbb{P}_g[\mathcal{B}_x^{-}]\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t) P g [ B x − ] ≤ Φ (( μ + θ ( μ + κ )) t ) . Recall that, since r x ∈ E rx\in E r x ∈ E , the event B x − \mathcal{B}_x^{-} B x − depends only on the Gaussian projection onto U x ∩ span ( v r , v x ) ⊥ \mathcal{U}_x\cap\operatorname{span}(v_r,v_x)^\perp U x ∩ span ( v r , v x ) ⊥ . Let L ( x ) \mathcal{L}^{(x)} L ( x ) be the event that L \mathcal{L} L occurs and x x x is the maximizer of M F j ( r ) M_{F_j}(r) M F j ( r ) (ties are broken by the fixed total order on V V V ). By what we have shown above for B x \mathcal{B}_x B x , applied to − g -g − g , if L ( x ) \mathcal{L}^{(x)} L ( x ) occurs and − ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t -\langle g,u_y\rangle\le q_{A'}(\rho_y)t − ⟨ g , u y ⟩ ≤ q A ′ ( ρ y ) t for every y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) , then B x − \mathcal{B}_x^{-} B x − occurs.
Since v r x ∈ U r ⊆ v r ⊥ v_{rx}\in\mathcal{U}_r\subseteq v_r^\perp v r x ∈ U r ⊆ v r ⊥ and v r x ∈ span ( v r , v x ) v_{rx}\in\operatorname{span}(v_r,v_x) v r x ∈ span ( v r , v x ) , for every z ∈ N F j ( r ) z\in N_{F_j}(r) z ∈ N F j ( r ) , we may write
v r z = ⟨ v r z , v r x ⟩ v r x + h z , h z ∈ U r ∩ span ( v r , v x ) ⊥ . v_{rz}=\langle v_{rz},v_{rx}\rangle v_{rx}+h_z,\qquad h_z\in\mathcal{U}_r\cap\operatorname{span}(v_r,v_x)^\perp. v r z = ⟨ v r z , v r x ⟩ v r x + h z , h z ∈ U r ∩ span ( v r , v x ) ⊥ .
Thus, conditional on ⟨ g , v r x ⟩ = s \langle g,v_{rx}\rangle=s ⟨ g , v r x ⟩ = s , we have
⟨ g , v r z ⟩ = ⟨ v r z , v r x ⟩ s + ⟨ g , h z ⟩ , \langle g,v_{rz}\rangle=\langle v_{rz},v_{rx}\rangle s+\langle g,h_z\rangle, ⟨ g , v r z ⟩ = ⟨ v r z , v r x ⟩ s + ⟨ g , h z ⟩ ,
so L ( x ) \mathcal{L}^{(x)} L ( x ) depends only on the Gaussian projection onto U r ∩ span ( v r , v x ) ⊥ \mathcal{U}_r\cap\operatorname{span}(v_r,v_x)^\perp U r ∩ span ( v r , v x ) ⊥ . On the other hand, B x − \mathcal{B}_x^{-} B x − depends only on the Gaussian projection onto U x ∩ span ( v r , v x ) ⊥ \mathcal{U}_x\cap\operatorname{span}(v_r,v_x)^\perp U x ∩ span ( v r , v x ) ⊥ . Since both subspaces are orthogonal to v r x v_{rx} v r x , the Gaussian projections onto them are jointly independent of ⟨ g , v r x ⟩ \langle g,v_{rx}\rangle ⟨ g , v r x ⟩ . Consequently, conditioning on ⟨ g , v r x ⟩ = s \langle g,v_{rx}\rangle=s ⟨ g , v r x ⟩ = s does not change their joint Gaussian distribution. By Theorem 41 , the correlation between these two subspaces is at most 1 2 \frac{1}{2} 2 1 . Thus, Theorem 16 gives P g [ L ( x ) ∩ B x − ∣ ⟨ g , v r x ⟩ = s ] ≤ P g [ L ( x ) ∣ ⟨ g , v r x ⟩ = s ] 1 / p 1 Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 \mathbb{P}_g[\mathcal{L}^{(x)}\cap\mathcal{B}_x^{-}\mid\langle g,v_{rx}\rangle=s]\le\mathbb{P}_g[\mathcal{L}^{(x)}\mid\langle g,v_{rx}\rangle=s]^{1/p_1}\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2} P g [ L ( x ) ∩ B x − ∣ ⟨ g , v r x ⟩ = s ] ≤ P g [ L ( x ) ∣ ⟨ g , v r x ⟩ = s ] 1/ p 1 Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 . Since L ( x ) ⊆ { ⟨ g , v r x ⟩ > A t } \mathcal{L}^{(x)}\subseteq\{\langle g,v_{rx}\rangle>At\} L ( x ) ⊆ {⟨ g , v r x ⟩ > A t } , integration and Hölder’s inequality give
P g [ L ( x ) ∩ B x − ] ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 ∫ A t ∞ P g [ L ( x ) ∣ ⟨ g , v r x ⟩ = s ] 1 / p 1 ϕ ( s ) d s ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 Φ ‾ ( A t ) 1 − 1 / p 1 P g [ L ( x ) ] 1 / p 1 . \begin{aligned}
\mathbb{P}_g[\mathcal{L}^{(x)}\cap\mathcal{B}_x^{-}]\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2}\int_{At}^{\infty}\mathbb{P}_g[\mathcal{L}^{(x)}\mid\langle g,v_{rx}\rangle=s]^{1/p_1}\phi(s)\,ds \\
&\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2}\overline{\Phi}(At)^{1-1/p_1}\mathbb{P}_g[\mathcal{L}^{(x)}]^{1/p_1}.
\end{aligned} P g [ L ( x ) ∩ B x − ] ≤ Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 ∫ A t ∞ P g [ L ( x ) ∣ ⟨ g , v r x ⟩ = s ] 1/ p 1 ϕ ( s ) d s ≤ Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 Φ ( A t ) 1 − 1/ p 1 P g [ L ( x ) ] 1/ p 1 .
Finally, the events L ( x ) \mathcal{L}^{(x)} L ( x ) are pairwise disjoint and ∑ x ∈ N F j ( r ) P g [ L ( x ) ] = P g [ L ] \sum_{x\in N_{F_j}(r)}\mathbb{P}_g[\mathcal{L}^{(x)}]=\mathbb{P}_g[\mathcal{L}] ∑ x ∈ N F j ( r ) P g [ L ( x ) ] = P g [ L ] . Hence, by the inclusion above and Hölder’s inequality once more, the probability in the second assertion is at most
∑ x ∈ N F j ( r ) P g [ L ( x ) ∩ B x − ] ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 Φ ‾ ( A t ) 1 − 1 / p 1 ∑ x ∈ N F j ( r ) P g [ L ( x ) ] 1 / p 1 ≤ ( ∣ N F j ( r ) ∣ Φ ‾ ( A t ) ) 1 − 1 / p 1 Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 1 / p 2 P g [ L ] 1 / p 1 . \begin{aligned}
\sum_{x\in N_{F_j}(r)}\mathbb{P}_g[\mathcal{L}^{(x)}\cap\mathcal{B}_x^{-}]
\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2}\overline{\Phi}(At)^{1-1/p_1}
\sum_{x\in N_{F_j}(r)}\mathbb{P}_g[\mathcal{L}^{(x)}]^{1/p_1} \\
&\le(|N_{F_j}(r)|\overline{\Phi}(At))^{1-1/p_1}\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{1/p_2}\mathbb{P}_g[\mathcal{L}]^{1/p_1}.
\end{aligned} x ∈ N F j ( r ) ∑ P g [ L ( x ) ∩ B x − ] ≤ Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 Φ ( A t ) 1 − 1/ p 1 x ∈ N F j ( r ) ∑ P g [ L ( x ) ] 1/ p 1 ≤ ( ∣ N F j ( r ) ∣ Φ ( A t ) ) 1 − 1/ p 1 Φ (( μ + θ ( μ + κ )) t ) 1/ p 2 P g [ L ] 1/ p 1 .
Proof of the Tail Bound
Proof of Theorem 38. We prove the two bounds by induction on i i i . We call the bounds on G i \mathcal{G}_i G i and L i \mathcal{L}_i L i the global bound and the local bound, respectively. We take t t t sufficiently large that ( m 0 + 2 ) ε t ≤ t − 4 Φ ‾ ( t / 2 ) (m_0+2)\varepsilon_t\le t^{-4}\overline{\Phi}(t/2) ( m 0 + 2 ) ε t ≤ t − 4 Φ ( t /2 ) . Since m 0 = O ( 1 / w ) = O ( t 2 / log t ) m_0=O(1/w)=O(t^2/\log t) m 0 = O ( 1/ w ) = O ( t 2 / log t ) and ℓ = Θ ( log n ) = Θ ( t 2 ) \ell=\Theta(\log n)=\Theta(t^2) ℓ = Θ ( log n ) = Θ ( t 2 ) , we have m 0 = o ( ℓ ) m_0=o(\ell) m 0 = o ( ℓ ) , and hence m 0 + 1 ≤ ℓ m_0+1\le\ell m 0 + 1 ≤ ℓ for sufficiently large n n n . Below, r ∈ V r\in V r ∈ V is an arbitrary vertex, and the induction hypothesis is used for every vertex.
The following estimate is used to sum the correlation-dependent failure probabilities over all vertices. For P ∈ P P\in\mathcal{P} P ∈ P , choose ρ P ∗ ∈ { ρ x : x ∈ S r , P } \rho_P^*\in\{\rho_x:x\in S_{r,P}\} ρ P ∗ ∈ { ρ x : x ∈ S r , P } minimizing C D a C_{D_a} C D a if S r , P ≠ ∅ S_{r,P}\ne\varnothing S r , P = ∅ , and ρ P ∗ ∈ P \rho_P^*\in P ρ P ∗ ∈ P arbitrarily otherwise. Every branch of C D ( ρ ) CD(\rho) C D ( ρ ) has D D D -derivative at least 1 1 1 , and hence D a − D = 3 η D_a-D=3\eta D a − D = 3 η gives C D a ( ρ ) − C D ( ρ ) ≥ 3 η C_{D_a}(\rho)-C_D(\rho)\ge3\eta C D a ( ρ ) − C D ( ρ ) ≥ 3 η . Since every x ∈ V x\in V x ∈ V lies in S r , P S_{r,P} S r , P for some P ∈ P P\in\mathcal{P} P ∈ P , we have, for sufficiently large t t t ,
∑ x ∈ V 2 exp ( − ( C D a ( ρ x ) + 1 4 ) t 2 2 ) ≤ 2 ∑ P ∈ P ∣ S r , P ∣ exp ( − ( C D a ( ρ P ∗ ) + 1 4 ) t 2 2 ) ≤ 2 ∑ P ∈ P exp ( − ( 1 4 + C D a ( ρ P ∗ ) − C D ( ρ P ∗ ) − η ) t 2 2 ) ≤ 2 ∣ P ∣ exp ( − ( 1 4 + 2 η ) t 2 2 ) ≤ ε t . (7) \begin{aligned}
\sum_{x\in V}2\exp\left(-\left(C_{D_a}(\rho_x)+\frac14\right)\frac{t^2}{2}\right)
&\le2\sum_{P\in\mathcal{P}}|S_{r,P}|\exp\left(-\left(C_{D_a}(\rho_P^*)+\frac14\right)\frac{t^2}{2}\right)\\
&\le2\sum_{P\in\mathcal{P}}\exp\left(-\left(\frac14+C_{D_a}(\rho_P^*)-C_D(\rho_P^*)-\eta\right)\frac{t^2}{2}\right)\\
&\le2|\mathcal{P}|\exp\left(-\left(\frac14+2\eta\right)\frac{t^2}{2}\right)\\
&\le\varepsilon_t . \tag*{(7)}
\end{aligned} x ∈ V ∑ 2 exp ( − ( C D a ( ρ x ) + 4 1 ) 2 t 2 ) ≤ 2 P ∈ P ∑ ∣ S r , P ∣ exp ( − ( C D a ( ρ P ∗ ) + 4 1 ) 2 t 2 ) ≤ 2 P ∈ P ∑ exp ( − ( 4 1 + C D a ( ρ P ∗ ) − C D ( ρ P ∗ ) − η ) 2 t 2 ) ≤ 2∣ P ∣ exp ( − ( 4 1 + 2 η ) 2 t 2 ) ≤ ε t . (7)
where the second inequality uses ∣ S r , P ∣ ≤ e ( C D ( ρ P ∗ ) + η ) t 2 / 2 |S_{r,P}|\le e^{(C_D(\rho_P^*)+\eta)t^2/2} ∣ S r , P ∣ ≤ e ( C D ( ρ P ∗ ) + η ) t 2 /2 , which is the hypothesis of Theorem 38, that is, the band bound of Theorem 37 after failed band rounding, the third inequality uses C D a ( ρ P ∗ ) − C D ( ρ P ∗ ) ≥ 3 η C_{D_a}(\rho_P^*)-C_D(\rho_P^*)\ge3\eta C D a ( ρ P ∗ ) − C D ( ρ P ∗ ) ≥ 3 η , and the last inequality uses ∣ P ∣ = O ( log n ) = O ( t 2 ) |\mathcal{P}|=O(\log n)=O(t^2) ∣ P ∣ = O ( log n ) = O ( t 2 ) .
Base case. For every fixed vertex x x x , since ∣ N F 0 ( x ) ∣ ≤ e D t 2 / 2 |N_{F_0}(x)|\le e^{Dt^2/2} ∣ N F 0 ( x ) ∣ ≤ e D t 2 /2 , A 0 2 = D a + 1 4 A_0^2=D_a+\frac14 A 0 2 = D a + 4 1 and D a − D = 3 η D_a-D=3\eta D a − D = 3 η , by a union bound,
P g [ M F 0 ( x ) > A 0 t ] ≤ e D t 2 / 2 Φ ‾ ( A 0 t ) ≤ e − ( A 0 2 − D ) t 2 / 2 = e − ( 1 / 4 + 3 η ) t 2 / 2 ≤ ε t . \mathbb{P}_g[M_{F_0}(x)>A_0t]\le e^{Dt^2/2}\overline{\Phi}(A_0t)\le e^{-(A_0^2-D)t^2/2}=e^{-(1/4+3\eta)t^2/2}\le\varepsilon_t. P g [ M F 0 ( x ) > A 0 t ] ≤ e D t 2 /2 Φ ( A 0 t ) ≤ e − ( A 0 2 − D ) t 2 /2 = e − ( 1/4 + 3 η ) t 2 /2 ≤ ε t .
Since F 1 ⊆ F 0 F_1\subseteq F_0 F 1 ⊆ F 0 , taking x = r x=r x = r gives P g [ L 0 ] ≤ ε t ≤ 2 ε t \mathbb{P}_g[\mathcal{L}_0]\le\varepsilon_t\le2\varepsilon_t P g [ L 0 ] ≤ ε t ≤ 2 ε t , which is the bound on L 0 \mathcal{L}_0 L 0 .
Since q c ( ρ x ) = ∥ u x ∥ C D a ( ρ x ) + 1 4 q_c(\rho_x)=\|u_x\|\sqrt{C_{D_a}(\rho_x)+\frac14} q c ( ρ x ) = ∥ u x ∥ C D a ( ρ x ) + 4 1 , a union bound, Theorem 12 and Equation (7) give
P g [ ∃ x ∈ F 0 : ∣ ⟨ g , u x ⟩ ∣ > q c ( ρ x ) t ] ≤ ∑ x ∈ V 2 exp ( − ( C D a ( ρ x ) + 1 4 ) t 2 2 ) ≤ ε t . \mathbb{P}_g\left[\exists x\in F_0:|\langle g,u_x\rangle|>q_c(\rho_x)t\right]\le\sum_{x\in V}2\exp\left(-\left(C_{D_a}(\rho_x)+\frac14\right)\frac{t^2}{2}\right)\le\varepsilon_t. P g [ ∃ x ∈ F 0 : ∣ ⟨ g , u x ⟩ ∣ > q c ( ρ x ) t ] ≤ x ∈ V ∑ 2 exp ( − ( C D a ( ρ x ) + 4 1 ) 2 t 2 ) ≤ ε t .
Since q A 0 = q c q_{A_0}=q_c q A 0 = q c by (P2) and F 1 ⊆ F 0 F_1\subseteq F_0 F 1 ⊆ F 0 , this gives P g [ G 0 ] ≤ ε t ≤ 2 ε t \mathbb{P}_g[\mathcal{G}_0]\le\varepsilon_t\le2\varepsilon_t P g [ G 0 ] ≤ ε t ≤ 2 ε t , which is the bound on G 0 \mathcal{G}_0 G 0 .
Induction step for the global bound. Assume that both bounds hold for some i < m 0 i<m_0 i < m 0 . For every fixed x ∈ F i + 2 x\in F_{i+2} x ∈ F i + 2 , applying the bound on L i \mathcal{L}_i L i with x x x in place of r r r gives
P g [ M F i + 1 ( x ) > A i t ] ≤ ( i + 2 ) ε t ≤ t − 4 Φ ‾ ( t / 2 ) . \mathbb{P}_g[M_{F_{i+1}}(x)>A_it]\le(i+2)\varepsilon_t\le t^{-4}\overline{\Phi}(t/2). P g [ M F i + 1 ( x ) > A i t ] ≤ ( i + 2 ) ε t ≤ t − 4 Φ ( t /2 ) .
For x ∈ F i + 2 x\in F_{i+2} x ∈ F i + 2 with f A i + 1 ( ρ x ) < q c ( ρ x ) f_{A_{i+1}}(\rho_x)<q_c(\rho_x) f A i + 1 ( ρ x ) < q c ( ρ x ) , take ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) of Case 1 with A = A i + 1 A=A_{i+1} A = A i + 1 , so that L = f A i + 1 ( ρ x ) 1 − ρ x 2 L=\frac{f_{A_{i+1}}(\rho_x)}{\sqrt{1-\rho_x^2}} L = 1 − ρ x 2 f A i + 1 ( ρ x ) . Since 0 ≤ A i − A i + 1 ≤ w 0\le A_i-A_{i+1}\le w 0 ≤ A i − A i + 1 ≤ w , the bound above and the assumption p F i + 1 ( x ) ≤ e − t 2 / 2 p_{F_{i+1}}(x)\le e^{-t^2/2} p F i + 1 ( x ) ≤ e − t 2 /2 for x ∈ F i + 2 x\in F_{i+2} x ∈ F i + 2 show that the assumptions of Theorem 42 hold with j = i + 2 j=i+2 j = i + 2 , A = A i + 1 A=A_{i+1} A = A i + 1 and A ′ = A i A'=A_i A ′ = A i . Hence (i) of Theorem 42 applies to such x x x , and (iii) of Theorem 40 gives L 2 + ( μ + θ ( μ + κ ) ) 2 ≥ C D a ( ρ x ) + 1 4 + 10 − 18 L^2+(\mu+\theta(\mu+\kappa))^2\ge C D_a(\rho_x)+\frac{1}{4}+10^{-18} L 2 + ( μ + θ ( μ + κ ) ) 2 ≥ C D a ( ρ x ) + 4 1 + 1 0 − 18 . Since q A i + 1 = min { q c , f A i + 1 } q_{A_{i+1}}=\min\{q_c,f_{A_{i+1}}\} q A i + 1 = min { q c , f A i + 1 } and f A f_A f A is increasing in A A A , we have
P g [ G i + 1 ] ≤ P g [ G i ] + P g [ G i + 1 ∩ G i c ] ≤ ( i + 2 ) ε t + ∑ x ∈ F i + 2 f A i + 1 ( ρ x ) < q c ( ρ x ) P g [ ∣ ⟨ g , u x ⟩ ∣ > f A i + 1 ( ρ x ) t , G i c ] ≤ ( i + 2 ) ε t + ∑ x ∈ V 2 exp ( − ( C D a ( ρ x ) + 1 4 ) t 2 2 ) ≤ ( i + 3 ) ε t , \begin{aligned}
\mathbb{P}_g[\mathcal{G}_{i+1}] &\le\mathbb{P}_g[\mathcal{G}_i]+\mathbb{P}_g[\mathcal{G}_{i+1}\cap\mathcal{G}_i^c]\\
&\le(i+2)\varepsilon_t+\sum_{\substack{x\in F_{i+2}\\f_{A_{i+1}}(\rho_x)<q_c(\rho_x)}}\mathbb{P}_g[|\langle g,u_x\rangle|>f_{A_{i+1}}(\rho_x)t,\mathcal{G}_i^c]\\
&\le(i+2)\varepsilon_t+\sum_{x\in V}2\exp\left(-\left(CD_a(\rho_x)+\frac{1}{4}\right)\frac{t^2}{2}\right)\\
&\le(i+3)\varepsilon_t,
\end{aligned} P g [ G i + 1 ] ≤ P g [ G i ] + P g [ G i + 1 ∩ G i c ] ≤ ( i + 2 ) ε t + x ∈ F i + 2 f A i + 1 ( ρ x ) < q c ( ρ x ) ∑ P g [ ∣ ⟨ g , u x ⟩ ∣ > f A i + 1 ( ρ x ) t , G i c ] ≤ ( i + 2 ) ε t + x ∈ V ∑ 2 exp ( − ( C D a ( ρ x ) + 4 1 ) 2 t 2 ) ≤ ( i + 3 ) ε t ,
where the second inequality uses the bound on G i \mathcal{G}_i G i , the third uses Theorem 42 , Theorem 12 and (iii) of Theorem 40, and the last uses Equation (7) ; this is the bound on G i + 1 \mathcal{G}_{i+1} G i + 1 .
3. Induction step for the local bound. Take ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) of Case 2 with A = A i + 1 A=A_{i+1} A = A i + 1 , so that L = A i + 1 L=A_{i+1} L = A i + 1 . Then (iii) of Theorem 40 gives A i + 1 2 + 3999 1000 ( μ + θ ( μ + κ ) ) 2 ≥ D a + 1 4 + 10 − 18 A_{i+1}^2+\frac{3999}{1000}(\mu+\theta(\mu+\kappa))^2\ge D_a+\frac{1}{4}+10^{-18} A i + 1 2 + 1000 3999 ( μ + θ ( μ + κ ) ) 2 ≥ D a + 4 1 + 1 0 − 18 . As in the global step, the assumptions of Theorem 42 hold with j = i + 2 j=i+2 j = i + 2 , A = A i + 1 A=A_{i+1} A = A i + 1 and A ′ = A i A'=A_i A ′ = A i , and the event L \mathcal{L} L in (ii) is L i + 1 \mathcal{L}_{i+1} L i + 1 . If L i + 1 \mathcal{L}_{i+1} L i + 1 occurs with the maximum attained by x x x and G i \mathcal{G}_i G i does not occur, then − ⟨ g , u y ⟩ ≤ q A i ( ρ y ) t -\langle g,u_y\rangle\le q_{A_i}(\rho_y)t − ⟨ g , u y ⟩ ≤ q A i ( ρ y ) t for every y ∈ N F i + 1 ( x ) y\in N_{F_{i+1}}(x) y ∈ N F i + 1 ( x ) ; that is, L i + 1 ∩ G i c \mathcal{L}_{i+1}\cap\mathcal{G}_i^c L i + 1 ∩ G i c is contained in the event in (ii) of Theorem 42 .
Then (ii) of Theorem 42 with p 1 = 11997 11996 p_1=\frac{11997}{11996} p 1 = 11996 11997 and p 2 = 3000 p_2=3000 p 2 = 3000 gives
P g [ L i + 1 ] ≤ P g [ G i ] + P g [ L i + 1 ∩ G i c ] ≤ ( i + 2 ) ε t + ( ∣ N F i + 2 ( r ) ∣ Φ ‾ ( A i + 1 t ) Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) 3999 / 1000 ) 1 / 11997 P g [ L i + 1 ] 11996 / 11997 ≤ ( i + 2 ) ε t + ( e D t 2 / 2 e − ( A i + 1 2 + 3999 1000 ( μ + θ ( μ + κ ) ) 2 ) t 2 / 2 ) 1 / 11997 P g [ L i + 1 ] 11996 / 11997 ≤ ( i + 2 ) ε t + ( ε t e − η t 2 ) 1 / 11997 P g [ L i + 1 ] 11996 / 11997 , \begin{aligned}
\mathbb{P}_g[\mathcal{L}_{i+1}] &\le\mathbb{P}_g[\mathcal{G}_i]+\mathbb{P}_g[\mathcal{L}_{i+1}\cap\mathcal{G}_i^c]\\
&\le(i+2)\varepsilon_t+\left(|N_{F_{i+2}}(r)|\overline{\Phi}(A_{i+1}t)\overline{\Phi}((\mu+\theta(\mu+\kappa))t)^{3999/1000}\right)^{1/11997}\mathbb{P}_g[\mathcal{L}_{i+1}]^{11996/11997}\\
&\le(i+2)\varepsilon_t+\left(e^{D t^2/2}e^{-(A_{i+1}^2+\frac{3999}{1000}(\mu+\theta(\mu+\kappa))^2)t^2/2}\right)^{1/11997}\mathbb{P}_g[\mathcal{L}_{i+1}]^{11996/11997}\\
&\le(i+2)\varepsilon_t+\left(\varepsilon_t e^{-\eta t^2}\right)^{1/11997}\mathbb{P}_g[\mathcal{L}_{i+1}]^{11996/11997},
\end{aligned} P g [ L i + 1 ] ≤ P g [ G i ] + P g [ L i + 1 ∩ G i c ] ≤ ( i + 2 ) ε t + ( ∣ N F i + 2 ( r ) ∣ Φ ( A i + 1 t ) Φ (( μ + θ ( μ + κ )) t ) 3999/1000 ) 1/11997 P g [ L i + 1 ] 11996/11997 ≤ ( i + 2 ) ε t + ( e D t 2 /2 e − ( A i + 1 2 + 1000 3999 ( μ + θ ( μ + κ ) ) 2 ) t 2 /2 ) 1/11997 P g [ L i + 1 ] 11996/11997 ≤ ( i + 2 ) ε t + ( ε t e − η t 2 ) 1/11997 P g [ L i + 1 ] 11996/11997 ,
where the second inequality uses the bound on G i \mathcal{G}_i G i and Theorem 42 , the third uses Theorem 12, and the last uses A i + 1 2 + 3999 1000 ( μ + θ ( μ + κ ) ) 2 ≥ D a + 1 4 A_{i+1}^2+\frac{3999}{1000}(\mu+\theta(\mu+\kappa))^2\ge D_a+\frac{1}{4} A i + 1 2 + 1000 3999 ( μ + θ ( μ + κ ) ) 2 ≥ D a + 4 1 and D a − D = 3 η D_a-D=3\eta D a − D = 3 η . Thus, it remains to show that the last term is at most ε t \varepsilon_t ε t .
Since x 11996 / 11997 ≤ 1 + x x^{11996/11997}\le1+x x 11996/11997 ≤ 1 + x for x ≥ 0 x\ge0 x ≥ 0 , applying this with x = P g [ L i + 1 ] / ε t x=\mathbb{P}_g[\mathcal{L}_{i+1}]/\varepsilon_t x = P g [ L i + 1 ] / ε t shows that the last term is at most e − η t 2 / 11997 ( ε t + P g [ L i + 1 ] ) e^{-\eta t^2/11997}(\varepsilon_t+\mathbb{P}_g[\mathcal{L}_{i+1}]) e − η t 2 /11997 ( ε t + P g [ L i + 1 ]) . Since e − η t 2 / 11997 ≤ 1 2 e^{-\eta t^2/11997}\le\frac{1}{2} e − η t 2 /11997 ≤ 2 1 for sufficiently large t t t , the display above then gives P g [ L i + 1 ] ≤ 2 ( i + 3 ) ε t ≤ 2 ( m 0 + 2 ) ε t \mathbb{P}_g[\mathcal{L}_{i+1}]\le2(i+3)\varepsilon_t\le2(m_0+2)\varepsilon_t P g [ L i + 1 ] ≤ 2 ( i + 3 ) ε t ≤ 2 ( m 0 + 2 ) ε t . Hence the last term is at most ( 2 m 0 + 5 ) e − η t 2 / 11997 ε t (2m_0+5)e^{-\eta t^2/11997}\varepsilon_t ( 2 m 0 + 5 ) e − η t 2 /11997 ε t , which is at most ε t \varepsilon_t ε t for sufficiently large t t t since m 0 = O ( t 2 / log t ) m_0=O(t^2/\log t) m 0 = O ( t 2 / log t ) . Therefore P g [ L i + 1 ] ≤ ( i + 3 ) ε t \mathbb{P}_g[\mathcal{L}_{i+1}]\le(i+3)\varepsilon_t P g [ L i + 1 ] ≤ ( i + 3 ) ε t , which is the bound on L i + 1 \mathcal{L}_{i+1} L i + 1 . This completes the induction. □
Improved Lifting to the Preceding Level
We now return to the Gaussian cover lifting argument. The convex set K 0 K_0 K 0 of small Gaussian measure is constructed from the local-maximum probability as before. For the convex set K 1 K_1 K 1 of large Gaussian measure, we use the prepared tail bound on the neighborhood maximum through Theorem 39 in place of the degree-based construction. This refinement allows us to maintain positive threshold drift for a larger target independent set. We keep the parameters and conventions of Subsection 5.4.
Lemma 43 (Improved single-step lifting). Let 1 ≤ j ≤ ℓ 1 \le j \le\ell 1 ≤ j ≤ ℓ , and let W ⊆ F j W \subseteq F_j W ⊆ F j be fixed independently of the Gaussian g g g . Suppose that { u x } x ∈ W \{u_x\}_{x \in W} { u x } x ∈ W is a ( U t , q ) (Ut,q) ( U t , q ) -cover for some 0 < U ≤ H 0 < U \le\sqrt{H} 0 < U ≤ H . Assume that p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x) \le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 and P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_g[M_{F_{j-1}}(x) > 1.968t] \le t^{-4}\overline{\Phi}(t/2) P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ( t /2 ) for every x ∈ W x \in W x ∈ W , and set W ′ = N F j − 1 ( W ) W' = N_{F_{j-1}}(W) W ′ = N F j − 1 ( W ) . Then, for sufficiently large t t t , { u y } y ∈ W ′ \{u_y\}_{y \in W'} { u y } y ∈ W ′ is a ( ( U + 1 / 3000 ) t , q − t − 100 ) ((U + 1/3000)t,q-t^{-100}) (( U + 1/3000 ) t , q − t − 100 ) -cover.
In place of the choices in Subsection 4.4, define ν = 437 2000 \nu= \frac{437}{2000} ν = 2000 437 , and let
ψ ( s ) = { 3 1 − s 2 , s ≤ ν , 3 ( 1 − ν s ) 1 − ν 2 , s ≥ ν . \psi(s) =
\begin{cases}
\sqrt{3}\sqrt{1-s^2}, & s \le\nu,\\
\frac{\sqrt{3}(1-\nu s)}{\sqrt{1-\nu^2}}, & s \ge\nu.
\end{cases} ψ ( s ) = { 3 1 − s 2 , 1 − ν 2 3 ( 1 − ν s ) , s ≤ ν , s ≥ ν .
Lemma 44. Let 1 ≤ j ≤ ℓ 1 \le j \le\ell 1 ≤ j ≤ ℓ , and let x ∈ F j x \in F_j x ∈ F j satisfy ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 , p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x) \le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 and P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_g[M_{F_{j-1}}(x) > 1.968t] \le t^{-4}\overline{\Phi}(t/2) P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ( t /2 ) . Let 0 < L < H 0 < L < \sqrt{H} 0 < L < H and ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] . Then, for sufficiently large t t t , over the Gaussian coordinates orthogonal to span ( v r , v x ) \operatorname{span}(v_r,v_x) span ( v r , v x ) ,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − L 2 ( 1 − ρ 2 ) ) − 200 w ) t ] ≥ Φ ( H − L 2 t ) . \mathbb{P}_g\left[\exists y \in N_{F_{j-1}}(x): \langle g,w_y\rangle+ L\rho\xi_y t \ge\left(\psi\left(\sqrt{H-L^2(1-\rho^2)}\right)-200w\right)t\right] \ge\Phi\left(\sqrt{H-L^2}t\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − L 2 ( 1 − ρ 2 ) ) − 200 w ) t ] ≥ Φ ( H − L 2 t ) .
Lemma 45. For every 0 ≤ U ≤ H 0 \le U \le\sqrt{H} 0 ≤ U ≤ H ,
3 2 ψ ( H − U 2 ) ≥ U 2 + 1 1500 . \frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^2}\right) \ge\frac{U}{2}+\frac{1}{1500}. 2 3 ψ ( H − U 2 ) ≥ 2 U + 1500 1 .
Theorem 44 is obtained by combining the bound on γ ( K 0 ( a , b ) ) \gamma(K_0(a,b)) γ ( K 0 ( a , b )) in Theorem 30 with the bound on γ ( K 1 ( A , c ) ) \gamma(K_1(A,c)) γ ( K 1 ( A , c )) in Theorem 39 through Theorem 29, and Theorem 45 turns the threshold ψ \psi ψ of Theorem 44 into a uniform positive drift. We first deduce Theorem 43 from Theorem 44 and Theorem 45.
Proof of Theorem 43. We follow the proof of Theorem 25, replacing the cover threshold ( U + 40 w ) t (U+40w)t ( U + 40 w ) t by ( U + 1 / 3000 ) t (U+1/3000)t ( U + 1/3000 ) t . Define E x \mathcal{E}_x E x and E x b a d \mathcal{E}_x^{\mathrm{bad}} E x bad as in that proof with this cover threshold.
Fix x ∈ W x \in W x ∈ W with ∥ u x ∥ > 0 \lVert u_x\rVert> 0 ∥ u x ∥ > 0 , and condition on Z x = − ⟨ g , u x / ∥ u x ∥ ⟩ = z Z_x=-\langle g,u_x/\lVert u_x\rVert\rangle=z Z x = − ⟨ g , u x / ∥ u x ∥⟩ = z , where U t / ∥ u x ∥ ≤ z < H t Ut/\lVert u_x\rVert\le z < \sqrt{H}t U t / ∥ u x ∥ ≤ z < H t . Set L = z / t L=z/t L = z / t , ρ = ρ x \rho=\rho_x ρ = ρ x , and U ′ = ∥ u x ∥ z / t U'=\lVert u_x\rVert z/t U ′ = ∥ u x ∥ z / t . Then U ≤ U ′ < H U \le U' < \sqrt{H} U ≤ U ′ < H . By Theorem 44,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − ( U ′ ) 2 ) − 200 w ) t | Z x = z ] ≥ Φ ( H t 2 − z 2 ) . \mathbb{P}_g\left[\exists y \in N_{F_{j-1}}(x): \langle g,w_y\rangle+L\rho\xi_y t \ge\left(\psi\left(\sqrt{H-(U')^2}\right)-200w\right)t \,\middle|\, Z_x=z\right] \ge\Phi\left(\sqrt{Ht^2-z^2}\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( H − ( U ′ ) 2 ) − 200 w ) t Z x = z ] ≥ Φ ( H t 2 − z 2 ) .
Hence, for every y y y as in this event and sufficiently large t t t ,
⟨ g , u y ⟩ ≥ U ′ t 2 + 3 2 ( ψ ( H − ( U ′ ) 2 ) − 200 w ) t ≥ U ′ t 2 + ( U ′ 2 + 1 1500 ) t − 100 3 w t ( ∵ x r e f ‘ T h e o r e m 45 ‘ ) = ( U ′ + 1 1500 − 100 3 w ) t ≥ ( U + 1 3000 ) t . ( ∵ U ′ ≥ U , 100 3 w ≤ 1 / 3000 ) \begin{aligned}
\langle g,u_y\rangle\ge\frac{U't}{2}+\frac{\sqrt{3}}{2}\left(\psi\left(\sqrt{H-(U')^2}\right)-200w\right)t \\
&\ge\frac{U't}{2}+\left(\frac{U'}{2}+\frac{1}{1500}\right)t-100\sqrt{3}wt \qquad(\because\ {xref}`Theorem 45`) \\
&=\left(U'+\frac{1}{1500}-100\sqrt{3}w\right)t \\
&\ge\left(U+\frac{1}{3000}\right)t. \qquad(\because\ U' \ge U,\ 100\sqrt{3}w \le1/3000)
\end{aligned} ⟨ g , u y ⟩ ≥ 2 U ′ t + 2 3 ( ψ ( H − ( U ′ ) 2 ) − 200 w ) t ≥ 2 U ′ t + ( 2 U ′ + 1500 1 ) t − 100 3 w t ( ∵ x r e f ‘ T h eor e m 45‘ ) = ( U ′ + 1500 1 − 100 3 w ) t ≥ ( U + 3000 1 ) t . ( ∵ U ′ ≥ U , 100 3 w ≤ 1/3000 )
By Theorem 44, P g [ E x b a d ∣ Z x = z ] ≤ Φ ‾ ( H t 2 − z 2 ) \mathbb{P}_{g}[\mathcal{E}_{x}^{\mathrm{bad}} \mid Z_x=z]\leq\overline{\Phi}(\sqrt{H}t^2-z^2) P g [ E x bad ∣ Z x = z ] ≤ Φ ( H t 2 − z 2 ) . The same integration over z z z and union bound as in the proof of Theorem 25 give
P g [ ⋃ x ∈ W E x b a d ] ≤ ( 1 + H t 2 π ) e − ( H − H 0 ) t 2 / 2 ≤ t − 100 \mathbb{P}_{g}\left[\bigcup_{x\in W}\mathcal{E}_{x}^{\mathrm{bad}}\right]\leq\left(1+\frac{\sqrt{H}t}{\sqrt{2\pi}}\right)e^{-(H-H_0)t^2/2}\leq t^{-100} P g [ x ∈ W ⋃ E x bad ] ≤ ( 1 + 2 π H t ) e − ( H − H 0 ) t 2 /2 ≤ t − 100
for sufficiently large t t t , where H − H 0 = 3 η > 0 H-H_0=3\eta>0 H − H 0 = 3 η > 0 . Gaussian symmetry and the same event inclusion now show that { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( U + 1 / 3000 ) t , q − t − 100 ) ((U+1/3000)t,q-t^{-100}) (( U + 1/3000 ) t , q − t − 100 ) -cover. □ \square □
Next, we prove Theorem 45.
Proof of Theorem 45. The estimate in the proof of Theorem 27, before substituting ν 2 = ( D − 3 ) / ( 4 D ) \nu^2=(D-3)/(4D) ν 2 = ( D − 3 ) / ( 4 D ) , gives
3 2 ψ ( H − U 2 ) − U 2 ≥ 3 2 1 − ν 2 − H ( 1 4 + 9 ν 2 4 ( 1 − ν 2 ) ) = 3 2 1 − ν 2 − H ( 1 + 8 ν 2 ) 2 1 − ν 2 \begin{aligned}
\frac{\sqrt{3}}{2}\psi\left(\sqrt{H-U^2}\right)-\frac{U}{2}\geq\frac{3}{2\sqrt{1-\nu^2}}-\sqrt{H\left(\frac{1}{4}+\frac{9\nu^2}{4(1-\nu^2)}\right)} \\
&=\frac{3}{2\sqrt{1-\nu^2}}-\frac{\sqrt{H(1+8\nu^2)}}{2\sqrt{1-\nu^2}}
\end{aligned} 2 3 ψ ( H − U 2 ) − 2 U ≥ 2 1 − ν 2 3 − H ( 4 1 + 4 ( 1 − ν 2 ) 9 ν 2 ) = 2 1 − ν 2 3 − 2 1 − ν 2 H ( 1 + 8 ν 2 )
for every 0 ≤ U ≤ H 0\leq U\leq\sqrt{H} 0 ≤ U ≤ H . Since H < 6.50621 H<6.50621 H < 6.50621 and ν = 437 2000 \nu=\frac{437}{2000} ν = 2000 437 ,
9 − H ( 1 + 8 ν 2 ) > 9 − 6.50621 ⋅ 690969 500000 = 441058251 50000000000 > 8 1000 . 9-H(1+8\nu^2)>9-6.50621\cdot\frac{690969}{500000}=\frac{441058251}{50000000000}>\frac{8}{1000}. 9 − H ( 1 + 8 ν 2 ) > 9 − 6.50621 ⋅ 500000 690969 = 50000000000 441058251 > 1000 8 .
Thus H ( 1 + 8 ν 2 ) < 3 \sqrt{H(1+8\nu^2)}<3 H ( 1 + 8 ν 2 ) < 3 , and rationalizing gives
3 − H ( 1 + 8 ν 2 ) 2 1 − ν 2 = 9 − H ( 1 + 8 ν 2 ) 2 1 − ν 2 ( 3 + H ( 1 + 8 ν 2 ) ) > 8 / 1000 12 = 1 1500 , \frac{3-\sqrt{H(1+8\nu^2)}}{2\sqrt{1-\nu^2}}=\frac{9-H(1+8\nu^2)}{2\sqrt{1-\nu^2}\left(3+\sqrt{H(1+8\nu^2)}\right)}>\frac{8/1000}{12}=\frac{1}{1500}, 2 1 − ν 2 3 − H ( 1 + 8 ν 2 ) = 2 1 − ν 2 ( 3 + H ( 1 + 8 ν 2 ) ) 9 − H ( 1 + 8 ν 2 ) > 12 8/1000 = 1500 1 ,
proving the claim. □ \square □
Now, we can prove Theorem 44.
Proof of Theorem 44. Put s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) . We choose the parameters so that ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c |L\rho-(1+\theta)b|\leq\theta c ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c , ensuring that the terms dependent on ξ y \xi_y ξ y produced by extrapolation have a nonnegative sum for every neighbor. If s ≤ ν s\leq\nu s ≤ ν , choose
a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , c = κ = θ = 0. a_0=\sqrt{1-s^2},\qquad b=L\rho,\qquad\mu=\sqrt{H-L^2},\qquad c=\kappa=\theta=0. a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , c = κ = θ = 0.
If s ≥ ν s\geq\nu s ≥ ν , choose
a 0 = 1 − ν 2 , b = ν L ρ s , μ = ν H − L 2 s , a_0=\sqrt{1-\nu^2},\qquad b=\frac{\nu L\rho}{s},\qquad\mu=\frac{\nu\sqrt{H-L^2}}{s}, a 0 = 1 − ν 2 , b = s ν L ρ , μ = s ν H − L 2 ,
c = 499 ∣ L ρ ∣ 1000 s , κ = 499 H − L 2 1000 s , θ = s − ν ν + 499 / 1000 . c=\frac{499|L\rho|}{1000s},\qquad\kappa=\frac{499\sqrt{H-L^2}}{1000s},\qquad\theta=\frac{s-\nu}{\nu+499/1000}. c = 1000 s 499∣ L ρ ∣ , κ = 1000 s 499 H − L 2 , θ = ν + 499/1000 s − ν .
Since L 2 ρ 2 + H − L 2 = s 2 L^{2}\rho^{2}+H-L^{2}=s^{2} L 2 ρ 2 + H − L 2 = s 2 , both choices satisfy
a 0 2 + b 2 + μ 2 = 1 , c 2 + κ 2 ≤ ( 499 1000 ) 2 < 1 4 , a_{0}^{2}+b^{2}+\mu^{2}=1,\qquad c^{2}+\kappa^{2}\leq\left(\frac{499}{1000}\right)^{2}<\frac{1}{4}, a 0 2 + b 2 + μ 2 = 1 , c 2 + κ 2 ≤ ( 1000 499 ) 2 < 4 1 ,
μ + θ ( μ + κ ) = H − L 2 , ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c . \mu+\theta(\mu+\kappa)=\sqrt{H-L^{2}},\qquad\lvert L\rho-(1+\theta)b\rvert\leq\theta c. μ + θ ( μ + κ ) = H − L 2 , ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c .
Moreover, a 0 ≥ 1 − ν 2 ≥ 3 2 a_{0}\geq\sqrt{1-\nu^{2}}\geq\frac{\sqrt{3}}{2} a 0 ≥ 1 − ν 2 ≥ 2 3 , ∣ b ∣ , μ ≤ ν < 1 4 \lvert b\rvert,\mu\leq\nu<\frac{1}{4} ∣ b ∣ , μ ≤ ν < 4 1 , and 0 ≤ θ < 4 0\leq\theta<4 0 ≤ θ < 4 . For s ≥ ν s\geq\nu s ≥ ν , the bound θ < 4 \theta<4 θ < 4 is equivalent to s < 5 ν + 4 ⋅ 499 1000 s<5\nu+4\cdot\frac{499}{1000} s < 5 ν + 4 ⋅ 1000 499 , which follows from s ≤ H s\leq\sqrt{H} s ≤ H and ( 5 ν + 4 ⋅ 499 1000 ) 2 − H > ( 5 ν + 4 ⋅ 499 1000 ) 2 − 6.50621 = 12130489 4000000 > 0 (5\nu+4\cdot\frac{499}{1000})^{2}-H>(5\nu+4\cdot\frac{499}{1000})^{2}-6.50621=\frac{12130489}{4000000}>0 ( 5 ν + 4 ⋅ 1000 499 ) 2 − H > ( 5 ν + 4 ⋅ 1000 499 ) 2 − 6.50621 = 4000000 12130489 > 0 .
For s ≤ ν s\leq\nu s ≤ ν , ( 1 + θ ) 3 a 0 − 1.968 θ = ψ ( s ) (1+\theta)\sqrt{3}a_{0}-1.968\theta=\psi(s) ( 1 + θ ) 3 a 0 − 1.968 θ = ψ ( s ) . For s ≥ ν s\geq\nu s ≥ ν , the exact inequality
3 ( 1 + 499 1000 ν ) 2 − 1.968 2 ( 1 − ν 2 ) = 6941936163 4000000000000 > 0 3\left(1+\frac{499}{1000}\nu\right)^{2}-1.968^{2}(1-\nu^{2})=\frac{6941936163}{4000000000000}>0 3 ( 1 + 1000 499 ν ) 2 − 1.96 8 2 ( 1 − ν 2 ) = 4000000000000 6941936163 > 0
gives
1.968 < 3 ( 1 + 499 ν / 1000 ) 1 − ν 2 = 3 a 0 + 3 ν ( ν + 499 / 1000 ) 1 − ν 2 , 1.968<\frac{\sqrt{3}(1+499\nu/1000)}{\sqrt{1-\nu^{2}}}=\sqrt{3}a_{0}+\frac{\sqrt{3}\nu(\nu+499/1000)}{\sqrt{1-\nu^{2}}}, 1.968 < 1 − ν 2 3 ( 1 + 499 ν /1000 ) = 3 a 0 + 1 − ν 2 3 ν ( ν + 499/1000 ) ,
by taking square roots, dividing by 1 − ν 2 \sqrt{1-\nu^{2}} 1 − ν 2 and using a 0 = 1 − ν 2 a_{0}=\sqrt{1-\nu^{2}} a 0 = 1 − ν 2 . Hence, since θ ≥ 0 \theta\geq0 θ ≥ 0 and θ ( ν + 499 1000 ) = s − ν \theta(\nu+\frac{499}{1000})=s-\nu θ ( ν + 1000 499 ) = s − ν , we obtain
( 1 + θ ) 3 a 0 − 1.968 θ ≥ 3 a 0 − 3 ν ( ν + 499 / 1000 ) 1 − ν 2 θ = 3 ( 1 − ν s ) 1 − ν 2 = ψ ( s ) . (1+\theta)\sqrt{3}a_{0}-1.968\theta\geq\sqrt{3}a_{0}-\frac{\sqrt{3}\nu(\nu+499/1000)}{\sqrt{1-\nu^{2}}}\theta=\frac{\sqrt{3}(1-\nu s)}{\sqrt{1-\nu^{2}}}=\psi(s). ( 1 + θ ) 3 a 0 − 1.968 θ ≥ 3 a 0 − 1 − ν 2 3 ν ( ν + 499/1000 ) θ = 1 − ν 2 3 ( 1 − ν s ) = ψ ( s ) .
Set a = a 0 − 20 w a=a_{0}-20w a = a 0 − 20 w . As in the proof of Theorem 26, for sufficiently large t t t , we have a ≥ 0 a\geq0 a ≥ 0 and
1 − a 2 − b 2 = μ 2 + 40 a 0 σ w − 400 w 2 ≥ μ 2 + 30 w , 1-a^{2}-b^{2}=\mu^{2}+40a_{0}\sigma w-400w^{2}\geq\mu^{2}+30w, 1 − a 2 − b 2 = μ 2 + 40 a 0 σ w − 400 w 2 ≥ μ 2 + 30 w ,
and hence
2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ≤ 2 π e 2 t − 13 e − μ 2 t 2 / 2 ≤ Φ ‾ ( μ t ) . 2\pi t^{2}e^{-(1-a^{2}-b^{2})t^{2}/2+a+\lvert b\rvert+t-2}\leq2\pi e^{2}t^{-13}e^{-\mu^{2}t^{2}/2}\leq\overline{\Phi}(\mu t). 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ≤ 2 π e 2 t − 13 e − μ 2 t 2 /2 ≤ Φ ( μ t ) .
The last inequality follows from Theorem 12 and the uniform bound μ ≤ ν \mu\leq\nu μ ≤ ν .
Since a ≥ 0 a\geq0 a ≥ 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\leq e^{-t^{2}/2} p F j − 1 ( x ) ≤ e − t 2 /2 , Theorem 30 and the display above give γ ( K 0 ( a , b ) ) ≤ Φ ‾ ( μ t ) \gamma(K_{0}(a,b))\leq\overline{\Phi}(\mu t) γ ( K 0 ( a , b )) ≤ Φ ( μ t ) .
Next, since c 2 + κ 2 < 1 4 c^{2}+\kappa^{2}<\frac{1}{4} c 2 + κ 2 < 4 1 and P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_{g}[M_{F_{j-1}}(x)>1.968t]\leq t^{-4}\overline{\Phi}(t/2) P g [ M F j − 1 ( x ) > 1.968 t ] ≤ t − 4 Φ ( t /2 ) , Theorem 39 with F = F j − 1 F=F_{j-1} F = F j − 1 gives, for sufficiently large t t t , γ ( K 1 ( 1.968 , c ) ) ≥ Φ ‾ ( κ t ) \gamma(K_{1}(1.968,c))\geq\overline{\Phi}(\kappa t) γ ( K 1 ( 1.968 , c )) ≥ Φ ( κ t ) .
Since ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_{y}\rangle=\langle g_{\perp},w_{y}\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ , the sets K 0 K_{0} K 0 and K 1 K_{1} K 1 in Theorem 29 for Y = N F j − 1 ( x ) Y=N_{F_{j-1}}(x) Y = N F j − 1 ( x ) , z y = w y z_{y}=w_{y} z y = w y , A y = ( 3 a − b ξ y ) t − 1 t A_{y}=(\sqrt{3}a-b\xi_{y})t-\frac{1}{t} A y = ( 3 a − b ξ y ) t − t 1 and B y = ( 1.968 − c ∣ ξ y ∣ ) t B_{y}=(1.968-c\lvert\xi_{y}\rvert)t B y = ( 1.968 − c ∣ ξ y ∣) t are K 0 ( a , b ) K_{0}(a,b) K 0 ( a , b ) and K 1 ( 1.968 , c ) K_{1}(1.968,c) K 1 ( 1.968 , c ) . Hence Theorem 29 with μ 0 = μ t \mu_{0}=\mu t μ 0 = μ t and κ 0 = κ t \kappa_{0}=\kappa t κ 0 = κ t gives γ ( K − θ ) ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) \gamma(K_{-\theta})\leq\overline{\Phi}((\mu+\theta(\mu+\kappa))t) γ ( K − θ ) ≤ Φ (( μ + θ ( μ + κ )) t ) , and the complement of K − θ K_{-\theta} K − θ is the event below. Therefore, since μ + θ ( μ + κ ) = H − L 2 \mu+\theta(\mu+\kappa)=\sqrt{H-L^{2}} μ + θ ( μ + κ ) = H − L 2 ,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( 1.968 − c ∣ ξ y ∣ ) t − 1 + θ t ] ≥ Φ ( H − L 2 t ) . \mathbb{P}_{g}\left[\exists y\in N_{F_{j-1}}(x):\langle g_{\perp},w_{y}\rangle>(1+\theta)(\sqrt{3}a-b\xi_{y})t-\theta(1.968-c\lvert\xi_{y}\rvert)t-\frac{1+\theta}{t}\right]\geq\Phi\left(\sqrt{H-L^{2}}t\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( 1.968 − c ∣ ξ y ∣) t − t 1 + θ ] ≥ Φ ( H − L 2 t ) .
For every y y y as in this event, we obtain
⟨ g , w y ⟩ + L ρ ξ y t > ( ( 1 + θ ) 3 a 0 − 1.968 θ + ( L ρ − ( 1 + θ ) b ) ξ y + θ c ∣ ξ y ∣ ) t − 20 3 ( 1 + θ ) w t − 1 + θ t . \langle g,w_{y}\rangle+L\rho\xi_{y}t>\left((1+\theta)\sqrt{3}a_{0}-1.968\theta+(L\rho-(1+\theta)b)\xi_{y}+\theta c\lvert\xi_{y}\rvert\right)t-20\sqrt{3}(1+\theta)wt-\frac{1+\theta}{t}. ⟨ g , w y ⟩ + L ρ ξ y t > ( ( 1 + θ ) 3 a 0 − 1.968 θ + ( L ρ − ( 1 + θ ) b ) ξ y + θ c ∣ ξ y ∣ ) t − 20 3 ( 1 + θ ) w t − t 1 + θ .
Hence, since ( 1 + θ ) 3 a 0 − 1.968 θ ≥ ψ ( s ) (1+\theta)\sqrt{3}a_{0}-1.968\theta\geq\psi(s) ( 1 + θ ) 3 a 0 − 1.968 θ ≥ ψ ( s ) , for sufficiently large t t t ,
⟨ g , w y ⟩ + L ρ ξ y t > ( ( 1 + θ ) 3 a 0 − 1.968 θ ) t − 20 3 ( 1 + θ ) w t − 1 + θ t ( ∵ ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c ) ≥ ψ ( s ) t − ( 100 3 + 5 log t ) w t ( ∵ θ < 4 ) ≥ ( ψ ( s ) − 200 w ) t . \begin{aligned}
\langle g,w_{y}\rangle+L\rho\xi_{y}t>((1+\theta)\sqrt{3}a_{0}-1.968\theta)t-20\sqrt{3}(1+\theta)wt-\frac{1+\theta}{t}\qquad(\because\ \lvert L\rho-(1+\theta)b\rvert\leq\theta c) \\
&\geq\psi(s)t-\left(100\sqrt{3}+\frac{5}{\log t}\right)wt\qquad(\because\ \theta<4) \\
&\geq(\psi(s)-200w)t.
\end{aligned} ⟨ g , w y ⟩ + L ρ ξ y t > (( 1 + θ ) 3 a 0 − 1.968 θ ) t − 20 3 ( 1 + θ ) w t − t 1 + θ ( ∵ ∣ L ρ − ( 1 + θ ) b ∣ ≤ θ c ) ≥ ψ ( s ) t − ( 100 3 + log t 5 ) w t ( ∵ θ < 4 ) ≥ ( ψ ( s ) − 200 w ) t .
Thus
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( s ) − 200 w ) t ] ≥ P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( 1.968 − c ∣ ξ y ∣ ) t − 1 + θ t ] ≥ Φ ( H − L 2 t ) , \begin{aligned}
\mathbb{P}_{g}\left[\exists y \in N_{F_{j-1}}(x): \langle g,w_y\rangle+L\rho\xi_y t \geq(\psi(s)-200w)t\right] \\
&\geq\mathbb{P}_{g}\left[\exists y \in N_{F_{j-1}}(x): \langle g_{\perp},w_y\rangle> (1+\theta)(\sqrt{3}a-b\xi_y)t-\theta(1.968-c|\xi_y|)t-\frac{1+\theta}{t}\right] \\
&\geq\Phi\left(\sqrt{H-L^2}t\right),
\end{aligned} P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ ( s ) − 200 w ) t ] ≥ P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( 3 a − b ξ y ) t − θ ( 1.968 − c ∣ ξ y ∣ ) t − t 1 + θ ] ≥ Φ ( H − L 2 t ) ,
which is the claim since s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) .
Improved Iterated Lifting
We now complete the analysis of SparseProgress (Algorithm 3).
In each trial of SparseProgress, we amplify every call to VectorColoring inside a band-rounding procedure by Θ ( log n ) \Theta(\log n) Θ ( log n ) independent repetitions and return the largest independent set over these repetitions. By Theorem 2 , each repetition meets its guarantee with probability at least 3 4 \frac{3}{4} 4 3 , so if the hidden constant is large enough, an amplified call misses its guarantee with probability at most n − 3 n^{-3} n − 3 . Since one trial makes at most n ∣ P ∣ n|\mathcal{P}| n ∣ P ∣ such calls, all of them meet their guarantees with probability at least 1 − ∣ P ∣ / n 2 = 1 − o ( 1 ) 1-|\mathcal{P}|/n^2=1-o(1) 1 − ∣ P ∣/ n 2 = 1 − o ( 1 ) .
The preparatory induction proceeds toward deeper cores, whereas cover lifting proceeds in the opposite direction. Since the preparation uses only m 0 = o ( ℓ ) m_0=o(\ell) m 0 = o ( ℓ ) levels and a constant number of improved lifts suffices, every lift takes place within the cores where the prepared tail bound on the neighborhood maximum is available.
Lemma 46. Let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with Δ ( G ) ≤ n b \Delta(G)\leq n^b Δ ( G ) ≤ n b . For sufficiently large n n n , one trial of SparseProgress runs in polynomial time and returns an independent set I I I with ∣ I ∣ ≥ T |I|\geq T ∣ I ∣ ≥ T with probability at least 1 2 \frac{1}{2} 2 1 . Consequently, for any fixed constant C > 0 C>0 C > 0 , running at most ⌈ C log 2 n ⌉ \lceil C\log_2 n\rceil ⌈ C log 2 n ⌉ independent trials gives a polynomial-time algorithm that returns such a set with probability at least 1 − n − C 1-n^{-C} 1 − n − C .
Proof. Consider one trial of SparseProgress, and suppose that all amplified VectorColoring calls satisfy their guarantees and that all empirical estimates used by Algorithm 2 satisfy
∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ / 8. |\widehat{p}_j(x)-p_{F_{j-1}}(x)|\leq\delta/8. ∣ p j ( x ) − p F j − 1 ( x ) ∣ ≤ δ /8.
The latter event has probability at least 3 4 \frac{3}{4} 4 3 by Theorem 21.
Suppose for contradiction that this execution returns no independent set of size at least T T T . Then every band-rounding procedure returns FAIL, so Theorem 37 gives, for every vertex r r r , every P ∈ P P\in\mathcal{P} P ∈ P , and every ρ ∈ P \rho\in P ρ ∈ P ,
∣ S r , P ∣ ≤ exp ( ( c D ( ρ ) + η ) t 2 2 ) . |S_{r,P}|\leq\exp\left((c_D(\rho)+\eta)\frac{t^2}{2}\right). ∣ S r , P ∣ ≤ exp ( ( c D ( ρ ) + η ) 2 t 2 ) .
Algorithm 2 must also return FAIL. By Theorem 19,
∣ F ℓ ∣ > n − 2 ℓ T δ = n − o ( n ) , |F_\ell|>n-\frac{2\ell T}{\delta}=n-o(n), ∣ F ℓ ∣ > n − δ 2 ℓ T = n − o ( n ) ,
and hence F ℓ ≠ ∅ F_\ell\neq\varnothing F ℓ = ∅ for sufficiently large n n n . Moreover, Theorem 21 gives
p F j − 1 ( x ) < δ = e − t 2 / 2 ( ∀ j , ∀ x ∈ F j ) . p_{F_{j-1}}(x)<\delta=e^{-t^2/2}\qquad(\forall j,\ \forall x\in F_j). p F j − 1 ( x ) < δ = e − t 2 /2 ( ∀ j , ∀ x ∈ F j ) .
Thus the assumptions of Theorem 38 hold. Choose a root r ∈ F ℓ r \in F_\ell r ∈ F ℓ , and let W 0 W_0 W 0 and U 0 U_0 U 0 be as in Subsection 4.3. By Theorem 24, { u x } x ∈ W 0 \{u_x\}_{x\in W_0} { u x } x ∈ W 0 is a ( U 0 t , 1 − O ( t e − t 2 / 4 ) ) (U_0t,1-O(te^{-t^2/4})) ( U 0 t , 1 − O ( t e − t 2 /4 )) -cover.
Let m m m be the smallest integer such that U m > H U_m > \sqrt{H} U m > H . Since H = D a + 2 H=D_a+2 H = D a + 2 is constant, m = O ( 1 ) m=O(1) m = O ( 1 ) . Also m 0 = o ( ℓ ) m_0=o(\ell) m 0 = o ( ℓ ) by Theorem 38, so m + m 0 + 1 < ℓ m+m_0+1<\ell m + m 0 + 1 < ℓ for sufficiently large n n n . For 0 ≤ i < m 0\leq i<m 0 ≤ i < m , define W i + 1 = N F ℓ − i − 2 ( W i ) W_{i+1}=N_{F_{\ell-i-2}}(W_i) W i + 1 = N F ℓ − i − 2 ( W i ) . Then W i ⊆ F ℓ − i − 1 W_i\subseteq F_{\ell-i-1} W i ⊆ F ℓ − i − 1 , and for 0 ≤ i < m 0\leq i<m 0 ≤ i < m , since ℓ − i − 1 ≥ ℓ − m > m 0 + 1 \ell-i-1\geq\ell-m>m_0+1 ℓ − i − 1 ≥ ℓ − m > m 0 + 1 , we have F ℓ − i − 2 ⊆ F m 0 + 1 F_{\ell-i-2}\subseteq F_{m_0+1} F ℓ − i − 2 ⊆ F m 0 + 1 . Since A m 0 = 1.968 A_{m_0}=1.968 A m 0 = 1.968 , applying the bound on L m 0 \mathcal{L}_{m_0} L m 0 of Theorem 38 to the vertex x x x gives
P g [ M F ℓ − i − 2 ( x ) > 1.968 t ] ≤ P g [ M F m 0 + 1 ( x ) > A m 0 t ] ≤ ( m 0 + 2 ) ε t ≤ t − 4 Φ ‾ ( t / 2 ) \mathbb{P}_g[M_{F_{\ell-i-2}}(x)>1.968t]\leq\mathbb{P}_g[M_{F_{m_0+1}}(x)>A_{m_0}t]\leq(m_0+2)\varepsilon_t\leq t^{-4}\overline{\Phi}(t/2) P g [ M F ℓ − i − 2 ( x ) > 1.968 t ] ≤ P g [ M F m 0 + 1 ( x ) > A m 0 t ] ≤ ( m 0 + 2 ) ε t ≤ t − 4 Φ ( t /2 )
for every x ∈ W i x\in W_i x ∈ W i . Hence repeated application of Theorem 43 gives that { u x } x ∈ W i \{u_x\}_{x\in W_i} { u x } x ∈ W i is a ( U i t , 1 − O ( t e − t 2 / 4 ) − i t − 100 ) (U_it,1-O(te^{-t^2/4})-it^{-100}) ( U i t , 1 − O ( t e − t 2 /4 ) − i t − 100 ) -cover for every 0 ≤ i ≤ m 0\leq i\leq m 0 ≤ i ≤ m . Since m = O ( 1 ) m=O(1) m = O ( 1 ) , it follows that, for sufficiently large n n n ,
P g [ max x ∈ W m ⟨ g , u x ⟩ ≥ U m t ] ≥ 1 2 . \mathbb{P}_g\left[\max_{x\in W_m}\langle g,u_x\rangle\geq U_mt\right]\geq\frac{1}{2}. P g [ x ∈ W m max ⟨ g , u x ⟩ ≥ U m t ] ≥ 2 1 .
On the other hand, ∥ u x ∥ ≤ 1 \|u_x\|\leq1 ∥ u x ∥ ≤ 1 and ∣ W m ∣ ≤ n |W_m|\leq n ∣ W m ∣ ≤ n . Since n = e H 0 t 2 / 2 n=e^{H_0t^2/2} n = e H 0 t 2 /2 , we have
P g [ max x ∈ W m ⟨ g , u x ⟩ ≥ U m t ] ≤ n Φ ‾ ( U m t ) ≤ e ( H 0 − U m 2 ) t 2 / 2 . \mathbb{P}_g\left[\max_{x\in W_m}\langle g,u_x\rangle\geq U_mt\right]\leq n\overline{\Phi}(U_mt)\leq e^{(H_0-U_m^2)t^2/2}. P g [ x ∈ W m max ⟨ g , u x ⟩ ≥ U m t ] ≤ n Φ ( U m t ) ≤ e ( H 0 − U m 2 ) t 2 /2 .
Since U m 2 > H = D a + 2 U_m^2>H=D_a+2 U m 2 > H = D a + 2 and H − H 0 = D a − D = 3 η > 0 H-H_0=D_a-D=3\eta>0 H − H 0 = D a − D = 3 η > 0 , the right-hand side is o ( 1 ) o(1) o ( 1 ) , a contradiction. Therefore the trial returns an independent set of size at least T T T whenever the events above hold. By a union bound, these events hold with probability at least 3 4 − o ( 1 ) ≥ 1 2 \frac{3}{4}-o(1)\geq\frac{1}{2} 4 3 − o ( 1 ) ≥ 2 1 in each trial, for sufficiently large n n n . Each trial takes polynomial time. Run at most ⌈ C log 2 n ⌉ \lceil C\log_2 n\rceil ⌈ C log 2 n ⌉ trials with independent randomness. The failure probability is at most 2 − ⌈ C log 2 n ⌉ ≤ n − C 2^{-\lceil C\log_2 n\rceil}\leq n^{-C} 2 − ⌈ C l o g 2 n ⌉ ≤ n − C , and the total running time remains polynomial. □ \square □
We now prove the independent-set theorem stated in Subsection 1.3.
Theorem 5. Let β = 0.692601 \beta=0.692601 β = 0.692601 . Let G = ( V , E ) G=(V,E) G = ( V , E ) be a 3-colorable graph on n n n vertices with maximum degree at most n β n^\beta n β . For every fixed C > 0 C>0 C > 0 , there is a randomized polynomial-time algorithm that, with probability at least 1 − n − C 1-n^{-C} 1 − n − C , returns an independent set I I I with ∣ I ∣ ≥ n ( 1 + β ) / 2 − o ( 1 ) |I|\geq n^{(1+\beta)/2-o(1)} ∣ I ∣ ≥ n ( 1 + β ) /2 − o ( 1 ) .
Proof of Theorem 5. For a graph with maximum degree at most n β n^\beta n β , Theorem 46 gives an independent set of size ∣ I ∣ ≥ T = n δ 100 ℓ 4 = n 1 + β 2 − o ( 1 ) |I|\geq T=\frac{n^\delta}{100\ell^4}=n^{\frac{1+\beta}{2}-o(1)} ∣ I ∣ ≥ T = 100 ℓ 4 n δ = n 2 1 + β − o ( 1 ) with probability at least 1 − n − C 1-n^{-C} 1 − n − C , when n n n is sufficiently large. The “− o ( 1 ) -o(1) − o ( 1 ) ” part allows us to ignore the case of small n n n , and Theorem 5 follows. □ \square □
Conclusion
In this paper, we studied the polynomial-time graph coloring algorithms for 3-colorable graphs. In § 4 , we presented a randomized algorithm that colors any 3-colorable graph using O ( n 1 / 6 + ε ) O(n^{1/6+\varepsilon}) O ( n 1/6 + ε ) colors for any ε > 0 \varepsilon>0 ε > 0 , using the Gaussian local-maximum selection algorithm. This result also yielded near-tight bounds on the chromatic number of vector k k k -colorable graphs (Subsection 1.4). In § 5 , we further improved the result to O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) colors using band-rounding algorithms and additional SDP constraints. This improves the previous best result of O ( n 0.1954 ) O(n^{0.1954}) O ( n 0.1954 ) colors due to Bansal, Huang, and Lee (2026).
However, a large gap still remains between the upper and lower bounds. On the lower bound side, the known hardness results concern only constant numbers of colors. Indeed, [9 ] proves the NP-hardness of coloring 3-colorable graphs with 5 colors, and the claimed result of a very recent preprint [FMW26] implies the NP-hardness of coloring 3-colorable graphs with any constant number of colors. Even under the stronger claimed result, the remaining gap is substantial. Therefore, it is natural to pose the following conjecture.
Conjecture 47. For every fixed ε > 0 \varepsilon> 0 ε > 0 , there exists a polynomial-time algorithm to color every 3-colorable graph with O ( n ε ) O(n^{\varepsilon}) O ( n ε ) colors.
Relation to Recent Preprints
As mentioned in Table 1, we are aware of recent preprints [46 , 45 , 3 , 48 ] that claim to color 3-colorable graphs in polynomial time, using as few as O ( n 0.1653 ) O(n^{0.1653}) O ( n 0.1653 ) colors. However, our work was conducted independently of these preprints.
AI Usage Disclosure
The authors extensively used OpenAI’s large language models, GPT-6 Astra and GPT-6 Pro, during the proof exploration phase, in collaboration with the authors. Initially, GPT-6 Astra generated an O ( n 0.1736 ) O(n^{0.1736}) O ( n 0.1736 ) -coloring algorithm. Then, we studied the analysis and explicitly suggested an idea to GPT-6 Pro, which developed it into an O ( n 0.1712 ) O(n^{0.1712}) O ( n 0.1712 ) -coloring algorithm. Building on this approach, GPT-6 Astra developed an O ( n 1 / 6 + ε ) O(n^{1/6+\varepsilon}) O ( n 1/6 + ε ) -coloring algorithm.
Subsequently, the authors generalized the strict vector 3-coloring algorithm to ordinary vector k k k -coloring for an arbitrary real number k > 2 k > 2 k > 2 , and with these hints, GPT-6 Astra generated an O ( n 0.16 ) O(n^{0.16}) O ( n 0.16 ) -coloring algorithm. We iteratively simplified its analysis, and GPT-6 Astra improved the simplified algorithm. Finally, GPT-6 Astra generated the final O ( n 0.1537 ) O(n^{0.1537}) O ( n 0.1537 ) -color algorithm.
This paper is entirely written by the human authors. However, we used GPT-5.6 Sol, GPT-6 Sol, GPT-6 Astra, and Claude Opus 5.5 to polish the presentation of the paper. Finally, the authors take full responsibility for the correctness and content of the paper.
Extension to Ordinary Vector k k k -Colorings
We prove Theorem 2 by extending the proof of Theorem 3 to ordinary vector k k k -colorings for every fixed real number k > 2 k > 2 k > 2 .
Each lemma below marked Replaces states the replacement of one lemma or corollary of the main text; the remaining lemmas are auxiliary. The proof of Theorem 3 goes through verbatim once these replacements are made. Algorithm 1, Algorithm 2, Theorem 19 and Theorem 21 are unchanged together with their proofs; Theorem 20 and Theorem 32 keep their statements, and their proofs go through with the parameters and lemmas of this appendix.
Throughout this appendix, set λ = 1 / ( k − 1 ) ∈ ( 0 , 1 ) \lambda= 1/(k-1) \in(0,1) λ = 1/ ( k − 1 ) ∈ ( 0 , 1 ) and c = k / ( k − 2 ) c = \sqrt{k/(k-2)} c = k / ( k − 2 ) ; for k = 3 k=3 k = 3 we have λ = 1 2 \lambda= \frac{1}{2} λ = 2 1 and c = 3 c = \sqrt{3} c = 3 . Nothing in this appendix uses that k k k is an integer: every identity below is an algebraic identity in the real parameter k > 2 k > 2 k > 2 . Recall that an ordinary vector k k k -coloring of G = ( V , E ) G=(V,E) G = ( V , E ) is a family of unit vectors { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V with ⟨ v x , v y ⟩ ≤ − λ \langle v_x,v_y\rangle\le-\lambda ⟨ v x , v y ⟩ ≤ − λ for every x y ∈ E xy \in E x y ∈ E .
Edge Residuals
For every vertex x ∈ V x \in V x ∈ V and y ∈ N V ( x ) y \in N_V(x) y ∈ N V ( x ) , write λ x y = − ⟨ v x , v y ⟩ \lambda_{xy}=-\langle v_x,v_y\rangle λ x y = − ⟨ v x , v y ⟩ , so that λ x y ≥ λ \lambda_{xy}\ge\lambda λ x y ≥ λ , and redefine
v x y = c 1 + λ x y ( v y + λ x y v x ) . v_{xy}=\frac{c}{1+\lambda_{xy}}(v_y+\lambda_{xy}v_x). v x y = 1 + λ x y c ( v y + λ x y v x ) .
For k = 3 k=3 k = 3 and λ x y = 1 2 \lambda_{xy}=\frac{1}{2} λ x y = 2 1 , this agrees with the vector v x y v_{xy} v x y of Subsection 3.2. For a Gaussian vector g g g , write g = Z v x + g ′ g=Zv_x+g' g = Z v x + g ′ with Z = ⟨ g , v x ⟩ Z=\langle g,v_x\rangle Z = ⟨ g , v x ⟩ and g ′ ⊥ v x g'\perp v_x g ′ ⊥ v x . This normalization keeps the local-maximum condition in the common form c Z > ⟨ g ′ , v x y ⟩ cZ>\langle g',v_{xy}\rangle c Z > ⟨ g ′ , v x y ⟩ for every edge (Theorem 49 ), although ⟨ v x , v y ⟩ = − λ x y \langle v_x,v_y\rangle=-\lambda_{xy} ⟨ v x , v y ⟩ = − λ x y may vary from edge to edge. In exchange, v x y v_{xy} v x y is no longer a unit vector, but it satisfies ∥ v x y ∥ ≤ 1 \|v_{xy}\|\le1 ∥ v x y ∥ ≤ 1 as seen below. This upper bound suffices, since each ⟨ g ′ , v x y ⟩ \langle g',v_{xy}\rangle ⟨ g ′ , v x y ⟩ is then a centered Gaussian with variance at most 1.
Lemma 48. For every x ∈ V x \in V x ∈ V and y ∈ N V ( x ) y \in N_V(x) y ∈ N V ( x ) , we have v x y ⊥ v x v_{xy}\perp v_x v x y ⊥ v x , ∥ v x y ∥ ≤ 1 \|v_{xy}\|\le1 ∥ v x y ∥ ≤ 1 , and
v y = − λ x y v x + 1 + λ x y c v x y . v_y=-\lambda_{xy}v_x+\frac{1+\lambda_{xy}}{c}v_{xy}. v y = − λ x y v x + c 1 + λ x y v x y . Proof. From the definition of v x y v_{xy} v x y , we have v y = − λ x y v x + 1 + λ x y c v x y v_y=-\lambda_{xy}v_x+\frac{1+\lambda_{xy}}{c}v_{xy} v y = − λ x y v x + c 1 + λ x y v x y . Since ⟨ v y + λ x y v x , v x ⟩ = ⟨ v y , v x ⟩ + λ x y ∥ v x ∥ 2 = − λ x y + λ x y = 0 \langle v_y+\lambda_{xy}v_x,v_x\rangle=\langle v_y,v_x\rangle+\lambda_{xy}\|v_x\|^2=-\lambda_{xy}+\lambda_{xy}=0 ⟨ v y + λ x y v x , v x ⟩ = ⟨ v y , v x ⟩ + λ x y ∥ v x ∥ 2 = − λ x y + λ x y = 0 , we have v y + λ x y v x ⊥ v x v_y+\lambda_{xy}v_x\perp v_x v y + λ x y v x ⊥ v x .
Moreover, ∥ v y + λ x y v x ∥ 2 = 1 − λ x y 2 \|v_y+\lambda_{xy}v_x\|^2=1-\lambda_{xy}^2 ∥ v y + λ x y v x ∥ 2 = 1 − λ x y 2 , so
∥ v x y ∥ 2 = c 2 1 − λ x y 1 + λ x y ≤ c 2 1 − λ 1 + λ = 1 , \|v_{xy}\|^2=c^2\frac{1-\lambda_{xy}}{1+\lambda_{xy}}\le c^2\frac{1-\lambda}{1+\lambda}=1, ∥ v x y ∥ 2 = c 2 1 + λ x y 1 − λ x y ≤ c 2 1 + λ 1 − λ = 1 , because the middle expression is decreasing in λ x y \lambda_{xy} λ x y and λ x y ≥ λ \lambda_{xy}\ge\lambda λ x y ≥ λ , and c 2 = k k − 2 = 1 + λ 1 − λ c^2=\frac{k}{k-2}=\frac{1+\lambda}{1-\lambda} c 2 = k − 2 k = 1 − λ 1 + λ .
Local-Maximum Probabilities
Lemma 49 (Replaces Theorem 17). Let F ⊆ V F\subseteq V F ⊆ V and x ∈ F x\in F x ∈ F . Then, for every a ∈ R a\in\mathbb{R} a ∈ R ,
p F ( x ) = E g [ Φ ‾ ( M F ( x ) / c ) ] , p_F(x)=\mathbb{E}_g\left[\overline{\Phi}\left(M_F(x)/c\right)\right], p F ( x ) = E g [ Φ ( M F ( x ) / c ) ] , P g [ M F ( x ) ≤ a ] ≤ p F ( x ) Φ ‾ ( a / c ) . \mathbb{P}_g[M_F(x)\le a]\le\frac{p_F(x)}{\overline{\Phi}(a/c)}. P g [ M F ( x ) ≤ a ] ≤ Φ ( a / c ) p F ( x ) .
Proof. Write g = Z v x + g ′ g = Z v_x + g' g = Z v x + g ′ with Z ∼ N ( 0 , 1 ) Z \sim N(0,1) Z ∼ N ( 0 , 1 ) independent of g ′ ⊥ v x g' \perp v_x g ′ ⊥ v x . By Theorem 48 , ⟨ g , v x y ⟩ = ⟨ g ′ , v x y ⟩ \langle g, v_{xy}\rangle= \langle g', v_{xy}\rangle ⟨ g , v x y ⟩ = ⟨ g ′ , v x y ⟩ , and for every y ∈ N F ( x ) y \in N_F(x) y ∈ N F ( x ) ,
⟨ g , v x ⟩ > ⟨ g , v y ⟩ ⟺ Z > − λ x y Z + 1 + λ x y c ⟨ g ′ , v x y ⟩ \langle g, v_x\rangle> \langle g, v_y\rangle\Longleftrightarrow Z > -\lambda_{xy}Z + \frac{1+\lambda_{xy}}{c}\langle g',v_{xy}\rangle ⟨ g , v x ⟩ > ⟨ g , v y ⟩ ⟺ Z > − λ x y Z + c 1 + λ x y ⟨ g ′ , v x y ⟩
⟺ c Z > ⟨ g ′ , v x y ⟩ . \Longleftrightarrow cZ > \langle g',v_{xy}\rangle. ⟺ c Z > ⟨ g ′ , v x y ⟩ .
Thus, x x x is a local maximum exactly when c Z > M F ( x ) cZ > M_F(x) c Z > M F ( x ) . The rest of the proof of Theorem 17 applies with 3 \sqrt{3} 3 replaced by c c c .
We prove a generalization of Theorem 18, which handles the low-degree regime.
Lemma 50 (Replaces Theorem 18). Let 0 < β ≤ k − 2 k 0 < \beta\le\frac{k-2}{k} 0 < β ≤ k k − 2 , and suppose that every vertex of G G G has degree at most n β n^\beta n β . Then there is a randomized polynomial-time algorithm that, with probability at least 3 4 \frac{3}{4} 4 3 , returns an independent set I I I with ∣ I ∣ ≥ n 1 − k − 2 k β − o ( 1 ) |I| \ge n^{1-\frac{k-2}{k}\beta-o(1)} ∣ I ∣ ≥ n 1 − k k − 2 β − o ( 1 ) .
Proof. Let a = 2 log ( 2 n β ) a = \sqrt{2\log(2n^\beta)} a = 2 log ( 2 n β ) and p ∗ = 1 2 Φ ‾ ( a / c ) p_* = \frac{1}{2}\overline{\Phi}(a/c) p ∗ = 2 1 Φ ( a / c ) , and run one trial of Algorithm 2 with δ = 1 2 \delta= \frac{1}{2} δ = 2 1 , T = ⌊ n p ∗ / 2 ⌋ T = \lfloor np_*/2\rfloor T = ⌊ n p ∗ /2 ⌋ and R = ⌈ 2 log 4 / p ∗ ⌉ R = \lceil2\log4/p_*\rceil R = ⌈ 2 log 4/ p ∗ ⌉ , as in the proof of Theorem 18. Fix x ∈ V x \in V x ∈ V . For each y ∈ N V ( x ) y \in N_V(x) y ∈ N V ( x ) , we have ∥ v x y ∥ ≤ 1 \lVert v_{xy}\rVert\le1 ∥ v x y ∥ ≤ 1 by Theorem 48 , so P g [ ⟨ g , v x y ⟩ > a ] ≤ P g [ ⟨ g , v x y ⟩ > a ∥ v x y ∥ ] ≤ Φ ‾ ( a ) \mathbb{P}_g[\langle g,v_{xy}\rangle> a] \le\mathbb{P}_g[\langle g,v_{xy}\rangle> a\lVert v_{xy}\rVert] \le\overline{\Phi}(a) P g [⟨ g , v x y ⟩ > a ] ≤ P g [⟨ g , v x y ⟩ > a ∥ v x y ∥] ≤ Φ ( a ) . Since x x x has at most n β n^\beta n β neighbors, a union bound and the upper bound of Theorem 12 give
P g [ M V ( x ) > a ] ≤ n β Φ ‾ ( a ) ≤ n β e − a 2 / 2 = 1 2 . \mathbb{P}_g[M_V(x)>a] \le n^\beta\overline{\Phi}(a) \le n^\beta e^{-a^2/2} = \frac{1}{2}. P g [ M V ( x ) > a ] ≤ n β Φ ( a ) ≤ n β e − a 2 /2 = 2 1 .
Hence, by Theorem 49 with F = V F=V F = V , p V ( x ) ≥ P g [ M V ( x ) ≤ a ] Φ ‾ ( a / c ) ≥ p ∗ p_V(x) \ge\mathbb{P}_g[M_V(x)\le a]\overline{\Phi}(a/c) \ge p_* p V ( x ) ≥ P g [ M V ( x ) ≤ a ] Φ ( a / c ) ≥ p ∗ . Since a 2 / ( 2 c 2 ) = k − 2 k log ( 2 n β ) a^2/(2c^2) = \frac{k-2}{k}\log(2n^\beta) a 2 / ( 2 c 2 ) = k k − 2 log ( 2 n β ) , the lower bound of Theorem 12 gives p ∗ = 1 2 Φ ‾ ( a / c ) ≥ c 0 2 ( 1 + a / c ) 2 − k − 2 k n − k − 2 k β ≥ n − k − 2 k β − o ( 1 ) p_* = \frac{1}{2}\overline{\Phi}(a/c) \ge\frac{c_0}{2(1+a/c)}2^{-\frac{k-2}{k}}n^{-\frac{k-2}{k}\beta} \ge n^{-\frac{k-2}{k}\beta-o(1)} p ∗ = 2 1 Φ ( a / c ) ≥ 2 ( 1 + a / c ) c 0 2 − k k − 2 n − k k − 2 β ≥ n − k k − 2 β − o ( 1 ) . The rest of the proof is the same as in Theorem 18.
The High-Degree Regime
For the rest of this appendix, assume k − 2 k < β ≤ 1 \frac{k-2}{k}<\beta\le1 k k − 2 < β ≤ 1 . Table 4 lists the parameters of this subsection that differ from Table 2 ; all other parameters, in particular T T T , R R R , t t t , H 0 H_0 H 0 , D D D , w w w and H H H , are given by the formulas of Table 2 .
Table 4. Parameters of Appendix § 8 .§ 3 that differ from Table 2 .
Here C ℓ = C ℓ ( k , β ) C_\ell=C_\ell(k,\beta) C ℓ = C ℓ ( k , β ) is a sufficiently large constant fixed in the proof of Theorem 55. Since ( k − 2 ) β 2 + k β < 1 \frac{(k-2)\beta}{2+k\beta}<1 2 + k β ( k − 2 ) β < 1 and ℓ = ⌈ log n ⌉ \ell=\lceil\log n\rceil ℓ = ⌈ log n ⌉ , we have 2 ℓ / δ = 2 n ( k − 2 ) β 2 + k β ℓ C ℓ + 1 = o ( n ) 2\ell/\delta=2n^{\frac{(k-2)\beta}{2+k\beta}}\ell^{C_\ell+1}=o(n) 2 ℓ / δ = 2 n 2 + k β ( k − 2 ) β ℓ C ℓ + 1 = o ( n ) , so the proof of Theorem 20 applies.
Lemma 51 (Replaces Theorem 24). Suppose that p F ℓ − 1 ( r ) ≤ e − t 2 / 2 p_{F_{\ell-1}}(r) \le e^{-t^2/2} p F ℓ − 1 ( r ) ≤ e − t 2 /2 . Then { u y } y ∈ W 0 \{u_y\}_{y\in W_0} { u y } y ∈ W 0 is a ( U 0 t , 1 − O ( t e − t 2 / 4 ) ) (U_0t,1-O(te^{-t^2/4})) ( U 0 t , 1 − O ( t e − t 2 /4 )) -cover.
Proof. Applying Theorem 49 with F = F ℓ − 1 F=F_{\ell-1} F = F ℓ − 1 , x = r x=r x = r and a = c t / 2 a=ct/\sqrt{2} a = c t / 2 gives
P g [ M F ℓ − 1 ( r ) ≤ c t / 2 ] ≤ e − t 2 / 2 Φ ‾ ( t / 2 ) = O ( t e − t 2 / 4 ) . \mathbb{P}_{g}[M_{F_{\ell-1}}(r)\le ct/\sqrt{2}]\le\frac{e^{-t^2/2}}{\overline{\Phi}(t/\sqrt{2})}=O(te^{-t^2/4}). P g [ M F ℓ − 1 ( r ) ≤ c t / 2 ] ≤ Φ ( t / 2 ) e − t 2 /2 = O ( t e − t 2 /4 ) .
Hence, with probability at least 1 − O ( t e − t 2 / 4 ) 1-O(te^{-t^2/4}) 1 − O ( t e − t 2 /4 ) , some y ∈ W 0 y\in W_0 y ∈ W 0 satisfies ⟨ g , v r y ⟩ ≥ c t / 2 \langle g,v_{ry}\rangle\ge ct/\sqrt{2} ⟨ g , v r y ⟩ ≥ c t / 2 . By Theorem 48 with x = r x=r x = r and v r y ⊥ v r v_{ry}\perp v_r v r y ⊥ v r , we have u y = 1 + λ r y c v r y u_y=\frac{1+\lambda_{ry}}{c}v_{ry} u y = c 1 + λ r y v r y , and since λ r y ≥ λ \lambda_{ry}\ge\lambda λ r y ≥ λ , we get ⟨ g , u y ⟩ ≥ 1 + λ c ⋅ c t 2 = U 0 t \langle g,u_y\rangle\ge\frac{1+\lambda}{c}\cdot\frac{ct}{\sqrt{2}}=U_0t ⟨ g , u y ⟩ ≥ c 1 + λ ⋅ 2 c t = U 0 t . □
:::
The following lemma replaces the identity for ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ preceding Theorem 26. For x ∈ F j x\in F_j x ∈ F j , write ρ = ρ x \rho=\rho_x ρ = ρ x , so that ∥ u x ∥ = 1 − ρ 2 \lVert u_x\rVert=\sqrt{1-\rho^2} ∥ u x ∥ = 1 − ρ 2 , and assume ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 . For y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) , decompose v x y v_{xy} v x y as in Equation (4) , writing the unit vector v x r v_{xr} v x r of Subsection 3.2 explicitly, since for r x ∈ E r_x\in E r x ∈ E it may differ from the edge residual defined above:
v x y = ξ y v r − ρ v x ∥ u x ∥ + w y = ξ y ∥ u x ∥ v r − ξ y ρ u x ∥ u x ∥ + w y , v_{xy}=\xi_y\frac{v_r-\rho v_x}{\lVert u_x\rVert}+w_y=\xi_y\lVert u_x\rVert v_r-\frac{\xi_y\rho u_x}{\lVert u_x\rVert}+w_y, v x y = ξ y ∥ u x ∥ v r − ρ v x + w y = ξ y ∥ u x ∥ v r − ∥ u x ∥ ξ y ρ u x + w y ,
where w y ⊥ span ( v r , v x ) w_y\perp\operatorname{span}(v_r,v_x) w y ⊥ span ( v r , v x ) ; by Theorem 48 , ( ξ y ) 2 + ∥ w y ∥ 2 ≤ 1 (\xi_y)^2+\lVert w_y\rVert^2\le1 ( ξ y ) 2 + ∥ w y ∥ 2 ≤ 1 .
Lemma 52. Suppose that ⟨ g , u x / ∥ u x ∥ ⟩ = − L t \langle g,u_x/\lVert u_x\rVert\rangle=-Lt ⟨ g , u x / ∥ u x ∥⟩ = − L t , and write U = ∥ u x ∥ L U=\lVert u_x\rVert L U = ∥ u x ∥ L . Let B ≥ − c U B\ge-cU B ≥ − c U , and suppose that ⟨ g , w y ⟩ + L ρ ξ y t ≥ B t \langle g,w_y\rangle+L\rho\xi_y t\ge Bt ⟨ g , w y ⟩ + L ρ ξ y t ≥ B t . Then ⟨ g , u y ⟩ ≥ ( λ U + 1 + λ c B ) t \langle g,u_y\rangle\ge\left(\lambda U+\frac{1+\lambda}{c}B\right)t ⟨ g , u y ⟩ ≥ ( λ U + c 1 + λ B ) t .
Proof. Projecting v y = − λ x y v x + 1 + λ x y c v x y v_y=-\lambda_{xy}v_x+\frac{1+\lambda_{xy}}{c}v_{xy} v y = − λ x y v x + c 1 + λ x y v x y of Theorem 48 onto v x ⊥ v_x^\perp v x ⊥ , we have
u y = − λ x y u x + 1 + λ x y c ( − ξ y ρ u x ∥ u x ∥ + w y ) . u_y=-\lambda_{xy}u_x+\frac{1+\lambda_{xy}}{c}\left(-\frac{\xi_y\rho u_x}{\lVert u_x\rVert}+w_y\right). u y = − λ x y u x + c 1 + λ x y ( − ∥ u x ∥ ξ y ρ u x + w y ) . Taking the inner product with g g g and using ⟨ g , u x ⟩ = − ∥ u x ∥ L t = − U t \langle g,u_x\rangle=-\lVert u_x\rVert Lt=-Ut ⟨ g , u x ⟩ = − ∥ u x ∥ L t = − U t ,
⟨ g , u y ⟩ = λ x y U t + 1 + λ x y c ( ⟨ g , w y ⟩ + L ρ ξ y t ) ≥ ( λ x y U + 1 + λ x y c B ) t . \langle g,u_y\rangle=\lambda_{xy}Ut+\frac{1+\lambda_{xy}}{c}\left(\langle g,w_y\rangle+L\rho\xi_y t\right)\ge\left(\lambda_{xy}U+\frac{1+\lambda_{xy}}{c}B\right)t. ⟨ g , u y ⟩ = λ x y U t + c 1 + λ x y ( ⟨ g , w y ⟩ + L ρ ξ y t ) ≥ ( λ x y U + c 1 + λ x y B ) t . Finally,
λ x y U + 1 + λ x y c B = λ U + 1 + λ c B + ( λ x y − λ ) ( U + B c ) ≥ λ U + 1 + λ c B , \lambda_{xy}U+\frac{1+\lambda_{xy}}{c}B=\lambda U+\frac{1+\lambda}{c}B+(\lambda_{xy}-\lambda)\left(U+\frac{B}{c}\right)\ge\lambda U+\frac{1+\lambda}{c}B, λ x y U + c 1 + λ x y B = λ U + c 1 + λ B + ( λ x y − λ ) ( U + c B ) ≥ λ U + c 1 + λ B , since λ x y ≥ λ \lambda_{xy}\ge\lambda λ x y ≥ λ and B ≥ − c U B\ge-cU B ≥ − c U . □
For k = 3 k=3 k = 3 and λ x y = 1 2 \lambda_{xy}=\frac{1}{2} λ x y = 2 1 , the second display is the identity for ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ in Subsection 4.4. Thus, by Theorem 52, an edge with λ x y > λ \lambda_{xy}>\lambda λ x y > λ yields at least the lower bound on ⟨ g , u y ⟩ \langle g,u_y\rangle ⟨ g , u y ⟩ obtained for λ x y = λ \lambda_{xy}=\lambda λ x y = λ , so the remaining lemmas only need to treat the case λ x y = λ \lambda_{xy}=\lambda λ x y = λ .
With the parameters of Table 4 , we have t 2 = 2 ( k − 2 ) β 2 + k β log n + 2 C ℓ log ℓ t^2=\frac{2(k-2)\beta}{2+k\beta}\log n+2C_\ell\log\ell t 2 = 2 + k β 2 ( k − 2 ) β log n + 2 C ℓ log ℓ . Then, as n → ∞ n\to\infty n → ∞ ,
D → 2 + k β k − 2 > k k − 2 = c 2 , D\to\frac{2+k\beta}{k-2}>\frac{k}{k-2}=c^2, D → k − 2 2 + k β > k − 2 k = c 2 ,
H → 2 + k β ( k − 2 ) β > 1 , H\to\frac{2+k\beta}{(k-2)\beta}>1, H → ( k − 2 ) β 2 + k β > 1 ,
where the first inequality uses β > k − 2 k \beta> \frac{k-2}{k} β > k k − 2 . Hence, D > c 2 D > c^2 D > c 2 and H > 1 H > 1 H > 1 for sufficiently large n n n . Since D > c 2 D > c^2 D > c 2 , the parameter ν \nu ν of Table 4 is well defined; let
ψ k ( s ) = { c 1 − s 2 , s ≤ ν , c ( 1 − ν s ) 1 − ν 2 , s ≥ ν . \psi_k(s)=
\begin{cases}
c\sqrt{1-s^2}, & s\le\nu,\\
\frac{c(1-\nu s)}{\sqrt{1-\nu^2}}, & s\ge\nu.
\end{cases} ψ k ( s ) = { c 1 − s 2 , 1 − ν 2 c ( 1 − ν s ) , s ≤ ν , s ≥ ν .
Thus ψ k \psi_k ψ k agrees with s ↦ c 1 − s 2 s\mapsto c\sqrt{1-s^2} s ↦ c 1 − s 2 on [ 0 , ν ] [0,\nu] [ 0 , ν ] and with its tangent line at s = ν s=\nu s = ν on [ ν , ∞ ) [\nu,\infty) [ ν , ∞ ) . For k = 3 k=3 k = 3 , these are the definitions of ν \nu ν and ψ \psi ψ in Subsection 4.4. Since D D D tends to a constant larger than c 2 c^2 c 2 , we have ν = Ω ( 1 ) \nu=\Omega(1) ν = Ω ( 1 ) , and ν 2 < k − 2 2 ( k − 1 ) ≤ 1 2 \nu^2<\frac{k-2}{2(k-1)}\le\frac{1}{2} ν 2 < 2 ( k − 1 ) k − 2 ≤ 2 1 .
Lemma 53 (Replaces Theorem 31). Let 0 < L < H 0<L<\sqrt{H} 0 < L < H , ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] and s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) . For sufficiently large n n n , there exist a 0 ≥ 1 − ν 2 a_0\ge\sqrt{1-\nu^2} a 0 ≥ 1 − ν 2 , κ ≥ 0 \kappa\ge0 κ ≥ 0 , 0 ≤ θ ≤ θ ˉ 0\le\theta\le\bar{\theta} 0 ≤ θ ≤ θ ˉ , ∣ b ∣ ≤ ν |b|\le\nu ∣ b ∣ ≤ ν and 0 ≤ μ ≤ ν 0\le\mu\le\nu 0 ≤ μ ≤ ν such that
a 0 2 + b 2 + μ 2 = 1 , a_0^2+b^2+\mu^2=1, a 0 2 + b 2 + μ 2 = 1 , μ + θ ( μ + κ ) = H − L 2 , \mu+\theta(\mu+\kappa)=\sqrt{H-L^2}, μ + θ ( μ + κ ) = H − L 2 , and
( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ k ( s ) . (1+\theta)ca_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=\psi_k(s). ( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ψ k ( s ) .
Proof. Since s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) , we have L 2 ρ 2 + ( H − L 2 ) = s 2 L^2\rho^2+(H-L^2)=s^2 L 2 ρ 2 + ( H − L 2 ) = s 2 .
First, suppose that s ≤ ν s\le\nu s ≤ ν . Choose
a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , κ = θ = 0. a_0=\sqrt{1-s^2},\qquad b=L\rho,\qquad\mu=\sqrt{H-L^2},\qquad\kappa=\theta=0. a 0 = 1 − s 2 , b = L ρ , μ = H − L 2 , κ = θ = 0.
By L 2 ρ 2 + ( H − L 2 ) = s 2 L^2\rho^2+(H-L^2)=s^2 L 2 ρ 2 + ( H − L 2 ) = s 2 , we have a 0 2 + b 2 + μ 2 = 1 a_0^2+b^2+\mu^2=1 a 0 2 + b 2 + μ 2 = 1 , ∣ b ∣ , μ ≤ s ≤ ν |b|,\mu\le s\le\nu ∣ b ∣ , μ ≤ s ≤ ν and a 0 ≥ 1 − ν 2 a_0\ge\sqrt{1-\nu^2} a 0 ≥ 1 − ν 2 . Also, μ + θ ( μ + κ ) = μ = H − L 2 \mu+\theta(\mu+\kappa)=\mu=\sqrt{H-L^2} μ + θ ( μ + κ ) = μ = H − L 2 . Since θ = 0 \theta=0 θ = 0 and b = L ρ b=L\rho b = L ρ ,
( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = c 1 − s 2 = ψ k ( s ) . (1+\theta)ca_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}=c\sqrt{1-s^2}=\psi_k(s). ( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = c 1 − s 2 = ψ k ( s ) .
Now suppose that s ≥ ν s\ge\nu s ≥ ν . Choose
a 0 = 1 − ν 2 , b = ν L ρ s , μ = ν H − L 2 s , a_0=\sqrt{1-\nu^2},\qquad b=\frac{\nu L\rho}{s},\qquad\mu=\frac{\nu\sqrt{H-L^2}}{s}, a 0 = 1 − ν 2 , b = s ν L ρ , μ = s ν H − L 2 ,
κ = D μ , θ = s / ν − 1 D + 1 . \kappa=D\mu,\qquad\theta=\frac{s/\nu-1}{D+1}. κ = D μ , θ = D + 1 s / ν − 1 .
As in the proof of Theorem 31, a 0 2 + b 2 + μ 2 = 1 a_0^2+b^2+\mu^2=1 a 0 2 + b 2 + μ 2 = 1 , μ + θ ( μ + κ ) = H − L 2 \mu+\theta(\mu+\kappa)=\sqrt{H-L^2} μ + θ ( μ + κ ) = H − L 2 and ∣ b ∣ , μ ≤ ν |b|,\mu\le\nu ∣ b ∣ , μ ≤ ν . Since s ≤ H s\le\sqrt{H} s ≤ H , we also have 0 ≤ θ ≤ θ ˉ 0\le\theta\le\bar{\theta} 0 ≤ θ ≤ θ ˉ . We have D ( 1 − ν 2 ) = c 2 ( 1 + D ν 2 ) D(1-\nu^2)=c^2(1+D\nu^2) D ( 1 − ν 2 ) = c 2 ( 1 + D ν 2 ) , since multiplying by k − 2 k-2 k − 2 turns it into ( k − 2 ) D − k = 2 ( k − 1 ) D ν 2 (k-2)D-k=2(k-1)D\nu^2 ( k − 2 ) D − k = 2 ( k − 1 ) D ν 2 , which is the definition of ν \nu ν . Hence, since L ρ − ( 1 + θ ) b = D θ b L\rho-(1+\theta)b=D\theta b L ρ − ( 1 + θ ) b = D θ b , κ = D μ \kappa=D\mu κ = D μ and b 2 + μ 2 = ν 2 b^2+\mu^2=\nu^2 b 2 + μ 2 = ν 2 ,
( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = θ 2 D ( 1 + D ν 2 ) = θ 2 c 2 ( 1 + D ν 2 ) 2 1 − ν 2 . \begin{aligned}
(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)=\theta^2D(1+D\nu^2) \\
&=\theta^2\frac{c^2(1+D\nu^2)^2}{1-\nu^2}.
\end{aligned} ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = θ 2 D ( 1 + D ν 2 ) = θ 2 1 − ν 2 c 2 ( 1 + D ν 2 ) 2 .
Since moreover a 0 = 1 − ν 2 a_0=\sqrt{1-\nu^2} a 0 = 1 − ν 2 and 1 + ( D + 1 ) θ = s / ν 1+(D+1)\theta=s/\nu 1 + ( D + 1 ) θ = s / ν ,
( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ( 1 + θ ) c 1 − ν 2 − θ c ( 1 + D ν 2 ) 1 − ν 2 = c 1 − ν 2 ( ( 1 + θ ) ( 1 − ν 2 ) − θ ( 1 + D ν 2 ) ) = c 1 − ν 2 ( 1 − ν 2 ( 1 + ( D + 1 ) θ ) ) = c ( 1 − ν s ) 1 − ν 2 = ψ k ( s ) . \begin{aligned}
(1+\theta)ca_0-\sqrt{(L\rho-(1+\theta)b)^2+\theta^2(D+\kappa^2)}
&=(1+\theta)c\sqrt{1-\nu^2}-\theta\frac{c(1+D\nu^2)}{\sqrt{1-\nu^2}}\\
&=\frac{c}{\sqrt{1-\nu^2}}\bigl((1+\theta)(1-\nu^2)-\theta(1+D\nu^2)\bigr)\\
&=\frac{c}{\sqrt{1-\nu^2}}\bigl(1-\nu^2(1+(D+1)\theta)\bigr)\\
&=\frac{c(1-\nu s)}{\sqrt{1-\nu^2}}\\
&=\psi_k(s).
\end{aligned} ( 1 + θ ) c a 0 − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) = ( 1 + θ ) c 1 − ν 2 − θ 1 − ν 2 c ( 1 + D ν 2 ) = 1 − ν 2 c ( ( 1 + θ ) ( 1 − ν 2 ) − θ ( 1 + D ν 2 ) ) = 1 − ν 2 c ( 1 − ν 2 ( 1 + ( D + 1 ) θ ) ) = 1 − ν 2 c ( 1 − ν s ) = ψ k ( s ) .
Lemma 54 (Replaces Theorem 26). Let 1 ≤ j ≤ ℓ 1\le j\le\ell 1 ≤ j ≤ ℓ , let x ∈ F j x\in F_j x ∈ F j satisfy p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 , let 0 < L < H 0<L<\sqrt{H} 0 < L < H and ρ ∈ [ − 1 , 1 ] \rho\in[-1,1] ρ ∈ [ − 1 , 1 ] , and set s = H − L 2 ( 1 − ρ 2 ) s=\sqrt{H-L^2(1-\rho^2)} s = H − L 2 ( 1 − ρ 2 ) . Then, for sufficiently large t t t , over the Gaussian coordinates orthogonal to span ( v r , v x ) \operatorname{span}(v_r,v_x) span ( v r , v x ) ,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ k ( s ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w ) t ] ≥ Φ ‾ ( H − L 2 t ) . \mathbb{P}_{g}\left[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle+L\rho\xi_y t\geq\left(\psi_k(s)-(20c+1)(1+\bar{\theta})w\right)t\right]\geq\overline{\Phi}\left(\sqrt{H-L^2}\,t\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ ( ψ k ( s ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w ) t ] ≥ Φ ( H − L 2 t ) .
Proof. As in the proof of Theorem 49, Theorem 30 holds with 3 \sqrt{3} 3 replaced by c c c , since the local-maximum condition becomes
c Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩ cZ_1-\xi_yZ_2>\langle g_{\perp},w_y\rangle c Z 1 − ξ y Z 2 > ⟨ g ⊥ , w y ⟩
for all y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) .
Take the parameters of Theorem 53. Then a 0 ≥ 1 − ν 2 > 1 / 2 a_0\geq\sqrt{1-\nu^2}>1/\sqrt{2} a 0 ≥ 1 − ν 2 > 1/ 2 , ∣ b ∣ , μ ≤ ν < 1 |b|,\mu\leq\nu<1 ∣ b ∣ , μ ≤ ν < 1 , and 0 ≤ θ ≤ θ ˉ 0\leq\theta\leq\bar{\theta} 0 ≤ θ ≤ θ ˉ . Set a = a 0 − 20 w a=a_0-20w a = a 0 − 20 w . For sufficiently large t t t , we have a ≥ 0 a\geq0 a ≥ 0 and
1 − a 2 − b 2 = μ 2 + 40 a 0 w − 400 w 2 ≥ μ 2 + 28 w , 1-a^2-b^2=\mu^2+40a_0w-400w^2\geq\mu^2+28w, 1 − a 2 − b 2 = μ 2 + 40 a 0 w − 400 w 2 ≥ μ 2 + 28 w ,
and hence, since e − 14 w t 2 = t − 14 e^{-14wt^2}=t^{-14} e − 14 w t 2 = t − 14 ,
2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ≤ O ( t − 12 ) e − μ 2 t 2 / 2 ≤ Φ ‾ ( μ t ) , 2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}\leq O(t^{-12})e^{-\mu^2t^2/2}\leq\overline{\Phi}(\mu t), 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ≤ O ( t − 12 ) e − μ 2 t 2 /2 ≤ Φ ( μ t ) ,
where the last inequality uses the lower bound of Theorem 12 and μ < 1 \mu<1 μ < 1 . Set B y = D + κ 2 1 − ( ξ y ) 2 t B_y=\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}\,t B y = D + κ 2 1 − ( ξ y ) 2 t for y ∈ N F j − 1 ( x ) y\in N_{F_{j-1}}(x) y ∈ N F j − 1 ( x ) . For each y y y , since ⟨ g , w y ⟩ ∼ N ( 0 , ∥ w y ∥ 2 ) \langle g,w_y\rangle\sim N(0,\lVert w_y\rVert^2) ⟨ g , w y ⟩ ∼ N ( 0 , ∥ w y ∥ 2 ) and ∥ w y ∥ ≤ 1 − ( ξ y ) 2 \lVert w_y\rVert\leq\sqrt{1-(\xi_y)^2} ∥ w y ∥ ≤ 1 − ( ξ y ) 2 , we have
P g [ ⟨ g , w y ⟩ > B y ] ≤ Φ ‾ ( D + κ 2 t ) . \mathbb{P}_g[\langle g,w_y\rangle>B_y]\leq\overline{\Phi}\left(\sqrt{D+\kappa^2}\,t\right). P g [⟨ g , w y ⟩ > B y ] ≤ Φ ( D + κ 2 t ) .
Since ∣ N F j − 1 ( x ) ∣ ≤ n β = e D t 2 / 2 |N_{F_{j-1}}(x)|\leq n^\beta=e^{Dt^2/2} ∣ N F j − 1 ( x ) ∣ ≤ n β = e D t 2 /2 , a union bound gives
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ > B y ] ≤ e D t 2 / 2 Φ ‾ ( D + κ 2 t ) ≤ Φ ‾ ( κ t ) , ( ∵ c i t e t ‘ T h e o r e m 13 ‘ ) \begin{aligned}
\mathbb{P}_g[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle>B_y]
&\leq e^{Dt^2/2}\overline{\Phi}\left(\sqrt{D+\kappa^2}\,t\right)\\
&\leq\overline{\Phi}(\kappa t),
\end{aligned}
\qquad(\because\ {citet}`Theorem 13`) P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ > B y ] ≤ e D t 2 /2 Φ ( D + κ 2 t ) ≤ Φ ( κ t ) , ( ∵ c i t e t ‘ T h eor e m 13‘ )
that is, γ ( K 1 ) ≥ Φ ( κ t ) \gamma(K_1)\geq\Phi(\kappa t) γ ( K 1 ) ≥ Φ ( κ t ) for K 1 = { g : ⟨ g , w y ⟩ ≤ B y ∀ y ∈ N F j − 1 ( x ) } K_1=\{g:\langle g,w_y\rangle\leq B_y\ \forall y\in N_{F_{j-1}}(x)\} K 1 = { g : ⟨ g , w y ⟩ ≤ B y ∀ y ∈ N F j − 1 ( x )} . Moreover, since a ≥ 0 a\geq0 a ≥ 0 and p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\leq e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 , Theorem 30 with 3 \sqrt{3} 3 replaced by c c c and the bound 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 / 2 + a + ∣ b ∣ + t − 2 ≤ Φ ‾ ( μ t ) 2\pi t^2e^{-(1-a^2-b^2)t^2/2+a+|b|+t^{-2}}\leq\overline{\Phi}(\mu t) 2 π t 2 e − ( 1 − a 2 − b 2 ) t 2 /2 + a + ∣ b ∣ + t − 2 ≤ Φ ( μ t ) above give γ ( K 0 ( a , b ) ) ≤ Φ ‾ ( μ t ) \gamma(K_0(a,b))\leq\overline{\Phi}(\mu t) γ ( K 0 ( a , b )) ≤ Φ ( μ t ) for K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( c a − b ξ y ) t − 1 t ∀ y ∈ N F j − 1 ( x ) } K_0(a,b)=\{g:\langle g_{\perp},w_y\rangle\leq(ca-b\xi_y)t-\frac{1}{t}\ \forall y\in N_{F_{j-1}}(x)\} K 0 ( a , b ) = { g : ⟨ g ⊥ , w y ⟩ ≤ ( c a − b ξ y ) t − t 1 ∀ y ∈ N F j − 1 ( x )} . Hence, since ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_{\perp},w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ , Theorem 29 for Y = N F j − 1 ( x ) Y=N_{F_{j-1}}(x) Y = N F j − 1 ( x ) , z y = w y z_y=w_y z y = w y and the thresholds ( c a − b ξ y ) t − 1 t (ca-b\xi_y)t-\frac{1}{t} ( c a − b ξ y ) t − t 1 and B y B_y B y , with μ 0 = μ t \mu_0=\mu t μ 0 = μ t and κ 0 = κ t \kappa_0=\kappa t κ 0 = κ t , gives γ ( K − θ ) ≤ Φ ‾ ( ( μ + θ ( μ + κ ) ) t ) \gamma(K-\theta)\le\overline{\Phi}((\mu+\theta(\mu+\kappa))t) γ ( K − θ ) ≤ Φ (( μ + θ ( μ + κ )) t ) . Taking complements, since μ + θ ( μ + κ ) = H − L 2 \mu+\theta(\mu+\kappa)=\sqrt{H-L^2} μ + θ ( μ + κ ) = H − L 2 , we obtain
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( c a − b ξ y ) t − θ D + κ 2 1 − ( ξ y ) 2 t − 1 + θ t ] ≥ Φ ( H − L 2 t ) . \mathbb{P}_g\left[\exists y\in N_{F_{j-1}}(x):\langle g_{\perp},w_y\rangle>(1+\theta)(ca-b\xi_y)t-\theta\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}\,t-\frac{1+\theta}{t}\right]\ge\Phi\left(\sqrt{H-L^2}\,t\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g ⊥ , w y ⟩ > ( 1 + θ ) ( c a − b ξ y ) t − θ D + κ 2 1 − ( ξ y ) 2 t − t 1 + θ ] ≥ Φ ( H − L 2 t ) .
For every y y y as in this event, since ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ \langle g,w_y\rangle=\langle g_{\perp},w_y\rangle ⟨ g , w y ⟩ = ⟨ g ⊥ , w y ⟩ ,
⟨ g , w y ⟩ + L ρ ξ y t > ( 1 + θ ) c a + ( L ρ − ( 1 + θ ) b ) ξ y − θ D + κ 2 1 − ( ξ y ) 2 t − 1 + θ t ≥ ( ( 1 + θ ) c a − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − 1 + θ t = ( ψ k ( s ) − 20 c ( 1 + θ ) w ) t − 1 + θ t ≥ ( ψ k ( s ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w ) t . \begin{aligned}
\langle g,w_y\rangle+L\rho\xi_y t
&>(1+\theta)ca+\bigl(L\rho-(1+\theta)b\bigr)\xi_y-\theta\sqrt{D+\kappa^2}\sqrt{1-(\xi_y)^2}\,t-\frac{1+\theta}{t}\\
&\ge\left((1+\theta)ca-\sqrt{\bigl(L\rho-(1+\theta)b\bigr)^2+\theta^2(D+\kappa^2)}\right)t-\frac{1+\theta}{t}\\
&=\bigl(\psi_k(s)-20c(1+\theta)w\bigr)t-\frac{1+\theta}{t}\\
&\ge\bigl(\psi_k(s)-(20c+1)(1+\bar{\theta})w\bigr)t.
\end{aligned} ⟨ g , w y ⟩ + L ρ ξ y t > ( 1 + θ ) c a + ( L ρ − ( 1 + θ ) b ) ξ y − θ D + κ 2 1 − ( ξ y ) 2 t − t 1 + θ ≥ ( ( 1 + θ ) c a − ( L ρ − ( 1 + θ ) b ) 2 + θ 2 ( D + κ 2 ) ) t − t 1 + θ = ( ψ k ( s ) − 20 c ( 1 + θ ) w ) t − t 1 + θ ≥ ( ψ k ( s ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w ) t .
where the equality uses Theorem 53 and a = a 0 − 20 w a=a_0-20w a = a 0 − 20 w , and the last inequality uses θ ≤ θ ˉ \theta\le\bar{\theta} θ ≤ θ ˉ . This proves the claim.
Lemma 55 (Replaces Theorem 27). If C ℓ C_\ell C ℓ is sufficiently large, then, for sufficiently large n n n and every 0 ≤ U ≤ H 0\le U\le\sqrt{H} 0 ≤ U ≤ H ,
1 + λ c ψ k ( H − U 2 ) ≥ ( 1 − λ ) U + ( 1 + λ c ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 ) w . \frac{1+\lambda}{c}\psi_k\left(\sqrt{H-U^2}\right)\ge(1-\lambda)U+\left(\frac{1+\lambda}{c}(20c+1)(1+\bar{\theta})+40\right)w. c 1 + λ ψ k ( H − U 2 ) ≥ ( 1 − λ ) U + ( c 1 + λ ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 ) w . Proof. Put s = H − U 2 s=\sqrt{H-U^2} s = H − U 2 , so that U 2 + s 2 = H U^2+s^2=H U 2 + s 2 = H . First suppose that s ≥ ν s\ge\nu s ≥ ν . Then Cauchy–Schwarz gives
( 1 − λ ) U + ( 1 + λ ) ν 1 − ν 2 s ≤ H ( ( 1 − λ ) 2 + ( 1 + λ ) 2 ν 2 1 − ν 2 ) , (1-\lambda)U+\frac{(1+\lambda)\nu}{\sqrt{1-\nu^2}}s\le\sqrt{H\left((1-\lambda)^2+\frac{(1+\lambda)^2\nu^2}{1-\nu^2}\right)}, ( 1 − λ ) U + 1 − ν 2 ( 1 + λ ) ν s ≤ H ( ( 1 − λ ) 2 + 1 − ν 2 ( 1 + λ ) 2 ν 2 ) , and hence
1 + λ c ψ k ( s ) − ( 1 − λ ) U = 1 + λ 1 − ν 2 − ( ( 1 − λ ) U + ( 1 + λ ) ν 1 − ν 2 s ) ≥ 1 + λ 1 − ν 2 − H ( ( 1 − λ ) 2 + ( 1 + λ ) 2 ν 2 1 − ν 2 ) = 2 k D ( k − 1 ) ( D + 1 ) − 2 H ( ( k − 2 ) D − 2 ) ( k − 1 ) ( D + 1 ) . \begin{aligned}
\frac{1+\lambda}{c}\psi_k(s)-(1-\lambda)U
&=\frac{1+\lambda}{\sqrt{1-\nu^2}}-\left((1-\lambda)U+\frac{(1+\lambda)\nu}{\sqrt{1-\nu^2}}s\right)\\
&\ge\frac{1+\lambda}{\sqrt{1-\nu^2}}-\sqrt{H\left((1-\lambda)^2+\frac{(1+\lambda)^2\nu^2}{1-\nu^2}\right)}\\
&=\sqrt{\frac{2kD}{(k-1)(D+1)}}-\sqrt{\frac{2H((k-2)D-2)}{(k-1)(D+1)}}.
\end{aligned} c 1 + λ ψ k ( s ) − ( 1 − λ ) U = 1 − ν 2 1 + λ − ( ( 1 − λ ) U + 1 − ν 2 ( 1 + λ ) ν s ) ≥ 1 − ν 2 1 + λ − H ( ( 1 − λ ) 2 + 1 − ν 2 ( 1 + λ ) 2 ν 2 ) = ( k − 1 ) ( D + 1 ) 2 k D − ( k − 1 ) ( D + 1 ) 2 H (( k − 2 ) D − 2 ) . where the last equality uses 1 + λ = k k − 1 1+\lambda=\frac{k}{k-1} 1 + λ = k − 1 k , 1 − λ = k − 2 k − 1 1-\lambda=\frac{k-2}{k-1} 1 − λ = k − 1 k − 2 and 1 − ν 2 = k ( D + 1 ) 2 ( k − 1 ) D 1-\nu^2=\frac{k(D+1)}{2(k-1)D} 1 − ν 2 = 2 ( k − 1 ) D k ( D + 1 ) . Let ζ k \zeta_k ζ k denote the last expression. Now suppose that s ≤ ν s\le\nu s ≤ ν . Then the left-hand side is ( 1 + λ ) 1 − s 2 − ( 1 − λ ) H − s 2 (1+\lambda)\sqrt{1-s^2}-(1-\lambda)\sqrt{H-s^2} ( 1 + λ ) 1 − s 2 − ( 1 − λ ) H − s 2 , and its derivative in s s s ,
s ( 1 − λ H − s 2 − 1 + λ 1 − s 2 ) , s\left(\frac{1-\lambda}{\sqrt{H-s^2}}-\frac{1+\lambda}{\sqrt{1-s^2}}\right), s ( H − s 2 1 − λ − 1 − s 2 1 + λ ) ,
is nonpositive for 0 ≤ s ≤ ν 0 \le s \le\nu 0 ≤ s ≤ ν , since H − s 2 > 1 − s 2 > 0 \sqrt{H-s^2} > \sqrt{1-s^2} > 0 H − s 2 > 1 − s 2 > 0 by H > 1 H > 1 H > 1 and ν < 1 \nu< 1 ν < 1 , and 1 − λ < 1 + λ 1-\lambda< 1+\lambda 1 − λ < 1 + λ . Hence it is at least its value at s = ν s=\nu s = ν , which is at least ζ k \zeta_k ζ k by the case s ≥ ν s \ge\nu s ≥ ν .
It remains to show that ζ k / w \zeta_k/w ζ k / w is at least 1 + λ c ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 \frac{1+\lambda}{c}(20c+1)(1+\bar{\theta})+40 c 1 + λ ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 . By the definitions of D D D , H 0 H_0 H 0 and
t 2 = 2 ( k − 2 ) β 2 + k β log n + 2 C ℓ log ℓ . t^2 = \frac{2(k-2)\beta}{2+k\beta}\log n + 2C_\ell\log\ell. t 2 = 2 + k β 2 ( k − 2 ) β log n + 2 C ℓ log ℓ .
k D − H 0 ( ( k − 2 ) D − 2 ) = 2 log n t 4 ( ( 2 + k β ) t 2 − 2 ( k − 2 ) β log n ) = 4 C ℓ ( 2 + k β ) log n log ℓ t 4 . kD-H_0((k-2)D-2) = \frac{2\log n}{t^4}\left((2+k\beta)t^2-2(k-2)\beta\log n\right)
= \frac{4C_\ell(2+k\beta)\log n\log\ell}{t^4}. k D − H 0 (( k − 2 ) D − 2 ) = t 4 2 log n ( ( 2 + k β ) t 2 − 2 ( k − 2 ) β log n ) = t 4 4 C ℓ ( 2 + k β ) log n log ℓ .
Since H = H 0 + 40 w H=H_0+40w H = H 0 + 40 w and w = log t / t 2 w=\log t/t^2 w = log t / t 2 , this gives
k D − H ( ( k − 2 ) D − 2 ) w = 4 C ℓ ( 2 + k β ) log n t 2 log ℓ log t − 40 ( ( k − 2 ) D − 2 ) → 4 C ℓ ( 2 + k β ) 2 ( k − 2 ) β − 40 k β . \frac{kD-H((k-2)D-2)}{w}
=4C_\ell(2+k\beta)\frac{\log n}{t^2}\frac{\log\ell}{\log t}-40((k-2)D-2)
\to\frac{4C_\ell(2+k\beta)^2}{(k-2)\beta}-40k\beta. w k D − H (( k − 2 ) D − 2 ) = 4 C ℓ ( 2 + k β ) t 2 log n log t log ℓ − 40 (( k − 2 ) D − 2 ) → ( k − 2 ) β 4 C ℓ ( 2 + k β ) 2 − 40 k β .
where we used log n t 2 → 2 + k β 2 ( k − 2 ) β \frac{\log n}{t^2}\to\frac{2+k\beta}{2(k-2)\beta} t 2 l o g n → 2 ( k − 2 ) β 2 + k β , log ℓ log t → 2 \frac{\log\ell}{\log t}\to2 l o g t l o g ℓ → 2 and ( k − 2 ) D − 2 → k β (k-2)D-2\to k\beta ( k − 2 ) D − 2 → k β . Rationalizing ζ k \zeta_k ζ k , we have
ζ k = 2 ( k D − H ( ( k − 2 ) D − 2 ) ) ( k − 1 ) ( D + 1 ) ⋅ 1 2 k D ( k − 1 ) ( D + 1 ) + 2 H ( ( k − 2 ) D − 2 ) ( k − 1 ) ( D + 1 ) . \zeta_k =
\frac{2(kD-H((k-2)D-2))}{(k-1)(D+1)}
\cdot
\frac{1}{\sqrt{\frac{2kD}{(k-1)(D+1)}}+\sqrt{\frac{2H((k-2)D-2)}{(k-1)(D+1)}}}. ζ k = ( k − 1 ) ( D + 1 ) 2 ( k D − H (( k − 2 ) D − 2 )) ⋅ ( k − 1 ) ( D + 1 ) 2 k D + ( k − 1 ) ( D + 1 ) 2 H (( k − 2 ) D − 2 ) 1 .
Moreover, D + 1 → k ( 1 + β ) k − 2 D+1\to\frac{k(1+\beta)}{k-2} D + 1 → k − 2 k ( 1 + β ) , and both radicands in ζ k \zeta_k ζ k tend to 2 ( 2 + k β ) ( k − 1 ) ( 1 + β ) \frac{2(2+k\beta)}{(k-1)(1+\beta)} ( k − 1 ) ( 1 + β ) 2 ( 2 + k β ) . Therefore
ζ k w → ( k − 1 ) ( 1 + β ) 2 ( 2 + k β ) ( 4 C ℓ ( 2 + k β ) 2 k β ( k − 1 ) ( 1 + β ) − 40 ( k − 2 ) β ( k − 1 ) ( 1 + β ) ) . \frac{\zeta_k}{w}\to
\sqrt{\frac{(k-1)(1+\beta)}{2(2+k\beta)}
\left(
\frac{4C_\ell(2+k\beta)^2}{k\beta(k-1)(1+\beta)}
-\frac{40(k-2)\beta}{(k-1)(1+\beta)}
\right)}. w ζ k → 2 ( 2 + k β ) ( k − 1 ) ( 1 + β ) ( k β ( k − 1 ) ( 1 + β ) 4 C ℓ ( 2 + k β ) 2 − ( k − 1 ) ( 1 + β ) 40 ( k − 2 ) β ) .
which grows linearly in C ℓ C_\ell C ℓ with a positive coefficient depending only on k k k and β \beta β . On the other hand, 1 + λ c ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 \frac{1+\lambda}{c}(20c+1)(1+\bar{\theta})+40 c 1 + λ ( 20 c + 1 ) ( 1 + θ ˉ ) + 40 converges to a constant that does not depend on C ℓ C_\ell C ℓ , so the claim follows by taking C ℓ C_\ell C ℓ sufficiently large. □ \square □
Lemma 56 (Replaces Theorem 25). Let 1 ≤ j ≤ ℓ 1 \le j \le\ell 1 ≤ j ≤ ℓ , and let W ⊆ F j W \subseteq F_j W ⊆ F j be fixed independently of the Gaussian g g g , and suppose that { u x } x ∈ W \{u_x\}_{x\in W} { u x } x ∈ W is a ( U t , q ) (Ut,q) ( U t , q ) -cover for some 0 < U ≤ H 0<U\le\sqrt{H} 0 < U ≤ H and q ∈ [ 0 , 1 ] q\in[0,1] q ∈ [ 0 , 1 ] . Assume that p F j − 1 ( x ) ≤ e − t 2 / 2 p_{F_{j-1}}(x)\le e^{-t^2/2} p F j − 1 ( x ) ≤ e − t 2 /2 for every x ∈ W x\in W x ∈ W , and set W ′ = N F j − 1 ( W ) W'=N_{F_{j-1}}(W) W ′ = N F j − 1 ( W ) . Then, for sufficiently large t t t , { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( U + 40 w ) t , q − t − 18 ) ((U+40w)t,q-t^{-18}) (( U + 40 w ) t , q − t − 18 ) -cover.**
Proof. We follow the proof of Theorem 25, with E x \mathcal{E}_x E x and E x b a d \mathcal{E}^{\mathrm{bad}}_x E x bad defined as there. Fix x ∈ W x\in W x ∈ W with ∥ u x ∥ > 0 \lVert u_x\rVert>0 ∥ u x ∥ > 0 , and condition on Z x = − ⟨ g , u x / ∥ u x ∥ ⟩ = z Z_x=-\langle g,u_x/\lVert u_x\rVert\rangle=z Z x = − ⟨ g , u x / ∥ u x ∥⟩ = z , where U t / ∥ u x ∥ ≤ z < H t Ut/\lVert u_x\rVert\le z<\sqrt{H}t U t / ∥ u x ∥ ≤ z < H t . Set L = z / t L=z/t L = z / t , ρ = ρ x \rho=\rho_x ρ = ρ x and U ′ = ∥ u x ∥ z / t U'=\lVert u_x\rVert z/t U ′ = ∥ u x ∥ z / t . Then U ≤ U ′ < H U\le U'<\sqrt{H} U ≤ U ′ < H and ( U ′ ) 2 = L 2 ( 1 − ρ 2 ) (U')^2=L^2(1-\rho^2) ( U ′ ) 2 = L 2 ( 1 − ρ 2 ) . Put B = ψ k ( H − ( U ′ ) 2 ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w B=\psi_k(\sqrt{H-(U')^2})-(20c+1)(1+\bar{\theta})w B = ψ k ( H − ( U ′ ) 2 ) − ( 20 c + 1 ) ( 1 + θ ˉ ) w . By Theorem 54,
P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ B t ∣ Z x = z ] ≥ Φ ( H t 2 − z 2 ) . \mathbb{P}_g\left[\exists y\in N_{F_{j-1}}(x):\langle g,w_y\rangle+L\rho\xi_y t\ge Bt\mid Z_x=z\right]\ge\Phi\left(\sqrt{Ht^2-z^2}\right). P g [ ∃ y ∈ N F j − 1 ( x ) : ⟨ g , w y ⟩ + L ρ ξ y t ≥ B t ∣ Z x = z ] ≥ Φ ( H t 2 − z 2 ) . By Theorem 55 at U ′ U' U ′ , we have λ U ′ + 1 + λ c B ≥ U ′ + 40 w \lambda U'+\frac{1+\lambda}{c}B\ge U'+40w λ U ′ + c 1 + λ B ≥ U ′ + 40 w , and in particular 1 + λ c B ≥ ( 1 − λ ) U ′ + 40 w > 0 \frac{1+\lambda}{c}B\ge(1-\lambda)U'+40w>0 c 1 + λ B ≥ ( 1 − λ ) U ′ + 40 w > 0 , so B > 0 ≥ − c U ′ B>0\ge-cU' B > 0 ≥ − c U ′ . Hence, for every y y y as in this event, Theorem 52 gives
⟨ g , u y ⟩ ≥ ( λ U ′ + 1 + λ c B ) t ≥ ( U ′ + 40 w ) t ≥ ( U + 40 w ) t . \langle g,u_y\rangle\ge\left(\lambda U'+\frac{1+\lambda}{c}B\right)t\ge(U'+40w)t\ge(U+40w)t. ⟨ g , u y ⟩ ≥ ( λ U ′ + c 1 + λ B ) t ≥ ( U ′ + 40 w ) t ≥ ( U + 40 w ) t .
Therefore P g [ E x b a d ∣ Z x = z ] ≤ Φ ‾ ( H t 2 − z 2 ) \mathbb{P}_{g}[\mathcal{E}_{x}^{\mathrm{bad}} \mid Z_x=z] \le\overline{\Phi}(\sqrt{Ht^2-z^2}) P g [ E x bad ∣ Z x = z ] ≤ Φ ( H t 2 − z 2 ) , and the same integration over z z z and union bound as in the proof of Theorem 25 give
P g [ ⋃ x ∈ W E x b a d ] ≤ ( 1 + H t 2 π ) e − ( H − H 0 ) t 2 / 2 ≤ t − 18 \mathbb{P}_{g}\left[\bigcup_{x\in W}\mathcal{E}_{x}^{\mathrm{bad}}\right] \le\left(1+\frac{\sqrt{H}t}{\sqrt{2\pi}}\right)e^{-(H-H_0)t^2/2}\le t^{-18} P g [ x ∈ W ⋃ E x bad ] ≤ ( 1 + 2 π H t ) e − ( H − H 0 ) t 2 /2 ≤ t − 18
for sufficiently large t t t , where H − H 0 = 40 w H-H_0=40w H − H 0 = 40 w . Gaussian symmetry and the same event inclusion now show that { u y } y ∈ W ′ \{u_y\}_{y\in W'} { u y } y ∈ W ′ is a ( ( ( U + 40 w ) t , q − t − 18 ) (((U+40w)t,q-t^{-18}) ((( U + 40 w ) t , q − t − 18 ) -cover. □
Lemma 57 (Replaces Theorem 33 ). Let k − 2 k < β ≤ 1 \frac{k-2}{k}<\beta\le1 k k − 2 < β ≤ 1 , and suppose that every vertex of G G G has degree at most n β n^\beta n β . Then, there is a randomized polynomial-time algorithm that, with probability at least 3 4 \frac{3}{4} 4 3 , returns an independent set I I I with ∣ I ∣ ≥ n 1 − ( k − 2 ) β 2 + k β − o ( 1 ) |I|\ge n^{1-\frac{(k-2)\beta}{2+k\beta}-o(1)} ∣ I ∣ ≥ n 1 − 2 + k β ( k − 2 ) β − o ( 1 ) .
Proof. As in the proof of Theorem 33 , run one trial of Algorithm 2 on G G G and { v x } x ∈ V \{v_x\}_{x\in V} { v x } x ∈ V with the parameters of Table 4 . Since ℓ = O ( log n ) \ell=O(\log n) ℓ = O ( log n ) , we have δ − 1 = n ( k − 2 ) β 2 + k β polylog ( n ) \delta^{-1}=n^{\frac{(k-2)\beta}{2+k\beta}}\operatorname{polylog}(n) δ − 1 = n 2 + k β ( k − 2 ) β polylog ( n ) . One trial calls Algorithm 1 at most ℓ R \ell R ℓ R times, each call takes polynomial time, and R = O ( δ − 2 log n ) = n 2 ( k − 2 ) β 2 + k β + o ( 1 ) R=O(\delta^{-2}\log n)=n^{\frac{2(k-2)\beta}{2+k\beta}+o(1)} R = O ( δ − 2 log n ) = n 2 + k β 2 ( k − 2 ) β + o ( 1 ) , so one trial runs in polynomial time.
By Theorem 21, with probability at least 3 4 \frac{3}{4} 4 3 , every empirical estimate satisfies ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ / 8 |\hat p_j(x)-p_{F_{j-1}}(x)|\le\delta/8 ∣ p ^ j ( x ) − p F j − 1 ( x ) ∣ ≤ δ /8 . Conditioned on this event, Theorem 32 shows that Algorithm 2 returns SUCCESS(I I I ) for an independent set I I I with
∣ I ∣ ≥ T ≥ n δ 100 ℓ 4 = n 1 − ( k − 2 ) β 2 + k β − o ( 1 ) . |I|\ge T\ge\frac{n\delta}{100\ell^4}=n^{1-\frac{(k-2)\beta}{2+k\beta}-o(1)}. ∣ I ∣ ≥ T ≥ 100 ℓ 4 n δ = n 1 − 2 + k β ( k − 2 ) β − o ( 1 ) .
As noted at the beginning of this appendix, the proof of Theorem 32 goes through, with Theorem 51 and Theorem 56 in place of Theorem 24 and Theorem 25. □
Proof of Theorem 2 . Fix C > 0 C>0 C > 0 . Run ⌈ C log 4 n ⌉ \lceil C\log^4 n\rceil ⌈ C log 4 n ⌉ independent trials of Algorithm 2, using the parameters from Theorem 50 if 0 < β ≤ k − 2 k 0<\beta\le\frac{k-2}{k} 0 < β ≤ k k − 2 and from Theorem 57 if k − 2 k < β ≤ 1 \frac{k-2}{k}<\beta\le1 k k − 2 < β ≤ 1 . Return the largest independent set obtained. By Theorem 50 or Theorem 57, according to the regime, each trial returns an independent set of the required size with probability at least 3 4 \frac{3}{4} 4 3 . Hence the probability that no trial returns such a set is at most 4 − ⌈ C log 4 n ⌉ ≤ n − C 4^{-\lceil C\log^4 n\rceil}\le n^{-C} 4 − ⌈ C l o g 4 n ⌉ ≤ n − C . Since each trial runs in polynomial time and there are O ( log n ) O(\log n) O ( log n ) trials, the total running time remains polynomial. □
Remark 58. The constant C ℓ ( k , β ) C_\ell(k,\beta) C ℓ ( k , β ) , the limit of θ ˉ \bar\theta θ ˉ , the threshold on n n n and the o ( 1 ) o(1) o ( 1 ) terms in Theorem 2 depend on k k k and β \beta β , and they degenerate as k → 2 k\to2 k → 2 , where c → ∞ c\to\infty c → ∞ and H → ∞ H\to\infty H → ∞ , and as β → k − 2 k \beta\to\frac{k-2}{k} β → k k − 2 from above, where ν → 0 \nu\to0 ν → 0 . For fixed k > 2 k>2 k > 2 and fixed β \beta β this is harmless, but applying the theorem with a parameter k = k ( n ) k=k(n) k = k ( n ) tending to 2 would require tracking this dependence.
Polynomial-Time Implementation of BuildVector and Treatment of Approximate Constraints
In this appendix, we discuss the details of the SDP implementation and the treatment of its constraints. In Appendix § 9 .§ 1 , we describe BuildVector and prove Theorem 34. Next, in Appendix § 9 .§ 2 , we justify treating some of the approximate constraints of the SDP as exact.
Ellipsoid Method
Algorithm 6 BuildVector
Require: graph G = ( V , E ) G=(V,E) G = ( V , E )
Ensure: vectors { v i } i ∈ V \{v_i\}_{i\in V} { v i } i ∈ V and band witnesses S \mathcal{S} S as in Theorem 34
Run the ellipsoid method over X \mathcal{X} X with initial ball B ( I , R 1 ) B(I,R_1) B ( I , R 1 ) and oracle O \mathcal{O} O
Let ( X , S ) (X,\mathcal{S}) ( X , S ) be the output returned by O \mathcal{O} O
Take a Gram realization X i j = ⟨ v i , v j ⟩ X_{ij}=\langle v_i,v_j\rangle X ij = ⟨ v i , v j ⟩
return ( { v i } i ∈ V , S ) (\{v_i\}_{i\in V},\mathcal{S}) ({ v i } i ∈ V , S )
Here, let X = { X ∈ R V × V : X = X T , ∀ i ∈ V : X i i = 1 } \mathcal{X}=\{X\in\mathbb{R}^{V\times V}:X=X^{\mathsf{T}},\ \forall i\in V:X_{ii}=1\} X = { X ∈ R V × V : X = X T , ∀ i ∈ V : X ii = 1 } . We regard X \mathcal{X} X as an affine space of dimension n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 with the Frobenius norm, and write B ( X , R ) = { Y ∈ X : ∥ Y − X ∥ F ≤ R } B(X,R)=\{Y\in\mathcal{X}:\lVert Y-X\rVert_{\mathrm{F}}\le R\} B ( X , R ) = { Y ∈ X : ∥ Y − X ∥ F ≤ R } . We define K K K as the set of X ∈ X X\in\mathcal{X} X ∈ X satisfying
X i j ≥ − 1 2 for all i , j ∈ V , X_{ij}\ge-\frac{1}{2}\qquad\text{for all }i,j\in V, X ij ≥ − 2 1 for all i , j ∈ V ,
X i j ≤ − 1 2 + ε S D P for all i j ∈ E , X_{ij}\le-\frac{1}{2}+\varepsilon_{\mathrm{SDP}}\qquad\text{for all }ij\in E, X ij ≤ − 2 1 + ε SDP for all ij ∈ E ,
X i j + X i j ′ ≤ 1 2 for all j j ′ ∈ E and i ∈ V ∖ { j , j ′ } , X_{ij}+X_{ij'}\le\frac{1}{2}\qquad\text{for all }jj'\in E\text{ and }i\in V\setminus\{j,j'\}, X ij + X i j ′ ≤ 2 1 for all j j ′ ∈ E and i ∈ V ∖ { j , j ′ } ,
∑ j ∈ S 1 + 2 X i j 3 ≤ Θ 1 ( S ) + ε S D P 4 for all i ∈ V and S ⊆ V , \sum_{j\in S}\frac{1+2X_{ij}}{3}\le\Theta_1(S)+\frac{\varepsilon_{\mathrm{SDP}}}{4}\qquad\text{for all }i\in V\text{ and }S\subseteq V, j ∈ S ∑ 3 1 + 2 X ij ≤ Θ 1 ( S ) + 4 ε SDP for all i ∈ V and S ⊆ V ,
∑ j ∈ S 2 − 2 X i j 3 ≤ Θ 2 ( S ) + ε S D P 4 for all i ∈ V and S ⊆ V , \sum_{j\in S}\frac{2-2X_{ij}}{3}\le\Theta_2(S)+\frac{\varepsilon_{\mathrm{SDP}}}{4}\qquad\text{for all }i\in V\text{ and }S\subseteq V, j ∈ S ∑ 3 2 − 2 X ij ≤ Θ 2 ( S ) + 4 ε SDP for all i ∈ V and S ⊆ V ,
Γ r x ( X ) ⪰ − ε S D P I for all r x ∈ E . \Gamma_{rx}(X)\succeq-\varepsilon_{\mathrm{SDP}}I\qquad\text{for all }rx\in E. Γ r x ( X ) ⪰ − ε SDP I for all r x ∈ E .
Since X ⪰ 0 X\succeq0 X ⪰ 0 and the last family of constraints are semidefinite constraints on affine functions of X X X , and every other constraint is linear in X X X , the set K K K is convex. The ellipsoid method below does not need to decide whether a candidate matrix X X X lies in K K K . For each X X X , the oracle O \mathcal{O} O checks the band constraints only for the bands S r , P = { j ∈ V : X r j ∈ P } S_{r,P}=\{j\in V:X_{rj}\in P\} S r , P = { j ∈ V : X r j ∈ P } determined by X X X , rather than for all S ⊆ V S\subseteq V S ⊆ V , together with the polynomially many other constraints of K K K . If one of these constraints is violated, O \mathcal{O} O returns the corresponding halfspace, which contains K K K and excludes X X X ; otherwise, it outputs X X X with the band witnesses required by Theorem 34, although X X X need not lie in K K K . In particular, O \mathcal{O} O never cuts at a point of K K K .
Fix a proper 3-coloring of G G G , and let X ‾ ∈ X \overline{X}\in\mathcal{X} X ∈ X be given by X ‾ i j = 1 \overline{X}_{ij}=1 X ij = 1 if i i i and j j j have the same color and X ‾ i j = − 1 2 \overline{X}_{ij}=-\frac{1}{2} X ij = − 2 1 otherwise. For ε 0 = ε S D P 16 n \varepsilon_0=\frac{\varepsilon_{\mathrm{SDP}}}{16n} ε 0 = 16 n ε SDP , define X ~ = ( 1 − ε 0 ) X ‾ + ε 0 I \widetilde{X}=(1-\varepsilon_0)\overline{X}+\varepsilon_0I X = ( 1 − ε 0 ) X + ε 0 I . Let R 0 = ε 0 8 n R_0=\frac{\varepsilon_0}{8n} R 0 = 8 n ε 0 and R 1 = n R_1=n R 1 = n .
Algorithm 7 OutputOrCut oracle O \mathcal{O} O
Require: graph G = ( V , E ) G=(V,E) G = ( V , E ) , matrix X ∈ X X\in\mathcal{X} X ∈ X Ensure: ( X , S ) (X,\mathcal{S}) ( X , S ) , or a halfspace that contains K K K and excludes X X X
if X ⪰̸ 0 X\not\succeq0 X ⪰ 0 then
return { Y ∈ X : u T Y u ≥ 0 } \{Y\in\mathcal{X}:u^{\mathsf{T}}Yu\ge0\} { Y ∈ X : u T Y u ≥ 0 } for a vector u u u with u T X u < 0 u^{\mathsf{T}}Xu<0 u T X u < 0
if Γ r x ( X ) ⪰̸ − ε S D P I \Gamma_{rx}(X)\not\succeq-\varepsilon_{\mathrm{SDP}}I Γ r x ( X ) ⪰ − ε SDP I for some r x ∈ E rx\in E r x ∈ E then
return { Y ∈ X : u T Γ r x ( Y ) u ≥ − ε S D P ∥ u ∥ 2 } \{Y\in\mathcal{X}:u^{\mathsf{T}}\Gamma_{rx}(Y)u\ge-\varepsilon_{\mathrm{SDP}}\lVert u\rVert^{2}\} { Y ∈ X : u T Γ r x ( Y ) u ≥ − ε SDP ∥ u ∥ 2 } for a vector u u u with u T Γ r x ( X ) u < − ε S D P ∥ u ∥ 2 u^{\mathsf{T}}\Gamma_{rx}(X)u<-\varepsilon_{\mathrm{SDP}}\lVert u\rVert^{2} u T Γ r x ( X ) u < − ε SDP ∥ u ∥ 2
if X X X violates one of the second to fourth constraints of K K K then
return the halfspace given by the violated constraint
for r ∈ V r\in V r ∈ V and P ∈ P P\in\mathcal{P} P ∈ P do
S ← { j ∈ V : X r j ∈ P } S\leftarrow\{j\in V:X_{rj}\in P\} S ← { j ∈ V : X r j ∈ P }
( W , Θ ^ ) ← SolveTheta ( S , m ) (\mathcal{W},\widehat{\Theta})\leftarrow\operatorname{SolveTheta}(S,m) ( W , Θ ) ← SolveTheta ( S , m ) , where m = 1 m=1 m = 1 if P P P is positive and m = 2 m=2 m = 2 otherwise
if P P P is positive and ∑ j ∈ S 1 + 2 X r j 3 > Θ ^ + ε S D P 2 \sum_{j\in S}\frac{1+2X_{rj}}{3}>\widehat{\Theta}+\frac{\varepsilon_{\mathrm{SDP}}}{2} ∑ j ∈ S 3 1 + 2 X r j > Θ + 2 ε SDP then
return { Y ∈ X : ∑ j ∈ S 1 + 2 Y r j 3 ≤ Θ ^ + ε S D P 2 } \{Y\in\mathcal{X}:\sum_{j\in S}\frac{1+2Y_{rj}}{3}\le\widehat{\Theta}+\frac{\varepsilon_{\mathrm{SDP}}}{2}\} { Y ∈ X : ∑ j ∈ S 3 1 + 2 Y r j ≤ Θ + 2 ε SDP }
if P P P is negative and ∑ j ∈ S 2 − 2 X r j 3 > Θ ^ + ε S D P 2 \sum_{j\in S}\frac{2-2X_{rj}}{3}>\widehat{\Theta}+\frac{\varepsilon_{\mathrm{SDP}}}{2} ∑ j ∈ S 3 2 − 2 X r j > Θ + 2 ε SDP then
return { Y ∈ X : ∑ j ∈ S 2 − 2 Y r j 3 ≤ Θ ^ + ε S D P 2 } \{Y\in\mathcal{X}:\sum_{j\in S}\frac{2-2Y_{rj}}{3}\le\widehat{\Theta}+\frac{\varepsilon_{\mathrm{SDP}}}{2}\} { Y ∈ X : ∑ j ∈ S 3 2 − 2 Y r j ≤ Θ + 2 ε SDP }
( S r , P , W r , P ) ← ( S , W ) (S_{r,P},\mathcal{W}_{r,P})\leftarrow(S,\mathcal{W}) ( S r , P , W r , P ) ← ( S , W )
return ( X , S ) (X,\mathcal{S}) ( X , S ) , where S = { ( ρ P , S r , P , W r , P ) } r ∈ V , P ∈ P \mathcal{S}=\{(\rho_P,S_{r,P},\mathcal{W}_{r,P})\}_{r\in V,P\in\mathcal{P}} S = {( ρ P , S r , P , W r , P ) } r ∈ V , P ∈ P
Here, SolveTheta ( S , m ) \operatorname{SolveTheta}(S,m) SolveTheta ( S , m ) returns a feasible solution W \mathcal{W} W of Θ m ( S ) \Theta_m(S) Θ m ( S ) together with its objective value Θ ^ \widehat{\Theta} Θ such that Θ ^ ≥ Θ m ( S ) − ε S D P 4 \widehat{\Theta}\ge\Theta_m(S)-\frac{\varepsilon_{\mathrm{SDP}}}{4} Θ ≥ Θ m ( S ) − 4 ε SDP . The witness is W = ( x ∅ , { x j } j ∈ S ) \mathcal{W}=(x_{\varnothing},\{x_j\}_{j\in S}) W = ( x ∅ , { x j } j ∈ S ) for m = 1 m=1 m = 1 and W = ( x ∅ , { ( x j , 1 , x j , 2 ) } j ∈ S ) \mathcal{W}=(x_{\varnothing},\{(x_{j,1},x_{j,2})\}_{j\in S}) W = ( x ∅ , {( x j , 1 , x j , 2 ) } j ∈ S ) for m = 2 m=2 m = 2 . Since m ∈ { 1 , 2 } m\in\{1,2\} m ∈ { 1 , 2 } , the SDP defining Θ m ( S ) \Theta_m(S) Θ m ( S ) has polynomial size, and its feasible region is bounded and has an inverse-polynomial relative interior margin. Hence, standard polynomial-time algorithms [27 ] yield an exactly feasible witness with objective value at least Θ m ( S ) − ε S D P / 4 \Theta_m(S)-\varepsilon_{\mathrm{SDP}}/4 Θ m ( S ) − ε SDP /4 in poly ( n , log 1 / ε S D P ) \operatorname{poly}(n,\log1/\varepsilon_{\mathrm{SDP}}) poly ( n , log 1/ ε SDP ) time.
Lemma 59 (OutputOrCut oracle O \mathcal{O} O ). There is a polynomial-time oracle O \mathcal{O} O such that, given a matrix X ∈ X X\in\mathcal{X} X ∈ X , the following holds:
either O \mathcal{O} O returns a pair ( X , S ) (X,\mathcal{S}) ( X , S ) satisfying all the guarantees required in Theorem 34,
or O \mathcal{O} O returns a halfspace H ⊆ X \mathcal{H}\subseteq\mathcal{X} H ⊆ X such that K ⊆ H K\subseteq\mathcal{H} K ⊆ H and X ∉ H X\notin\mathcal{H} X ∈ / H .
Proof. We show that Algorithm 7 has the desired property. First, suppose that O \mathcal{O} O returns a halfspace H \mathcal{H} H . By construction, X ∉ H X\notin\mathcal{H} X ∈ / H . If H = { Y ∈ X : u T Y u ≥ 0 } \mathcal{H}=\{Y\in\mathcal{X}:u^{\mathsf{T}}Yu\ge0\} H = { Y ∈ X : u T Y u ≥ 0 } , then K ⊆ H K\subseteq\mathcal{H} K ⊆ H since every Y ∈ K Y\in K Y ∈ K satisfies Y ⪰ 0 Y\succeq0 Y ⪰ 0 . If H = { Y ∈ X : u T Γ r x ( Y ) u ≥ − ε S D P ∥ u ∥ 2 } \mathcal{H}=\{Y\in\mathcal{X}:u^{\mathsf{T}}\Gamma_{rx}(Y)u\ge-\varepsilon_{\mathrm{SDP}}\lVert u\rVert^{2}\} H = { Y ∈ X : u T Γ r x ( Y ) u ≥ − ε SDP ∥ u ∥ 2 } , then H \mathcal{H} H is a halfspace since Γ r x \Gamma_{rx} Γ r x is affine, and K ⊆ H K\subseteq\mathcal{H} K ⊆ H since every Y ∈ K Y\in K Y ∈ K satisfies Γ r x ( Y ) ⪰ − ε S D P I \Gamma_{rx}(Y)\succeq-\varepsilon_{\mathrm{SDP}}I Γ r x ( Y ) ⪰ − ε SDP I . If H \mathcal{H} H is given by one of the second to fourth constraints of K K K , then K ⊆ H K\subseteq\mathcal{H} K ⊆ H by the definition of K K K . Otherwise, H \mathcal{H} H is returned for some r ∈ V r\in V r ∈ V and P ∈ P P \in\mathcal{P} P ∈ P . If m = 1 m = 1 m = 1 , then H = { Y ∈ X : ∑ j ∈ S 1 + 2 Y r j 3 ≤ Θ ^ + ε S D P 2 } \mathcal{H} = \{Y \in\mathcal{X} : \sum_{j \in S} \frac{1+2Y_{rj}}{3} \le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2}\} H = { Y ∈ X : ∑ j ∈ S 3 1 + 2 Y r j ≤ Θ + 2 ε SDP } , and every Y ∈ K Y \in K Y ∈ K satisfies
∑ j ∈ S 1 + 2 Y r j 3 ≤ Θ 1 ( S ) + ε S D P 4 ( ∵ Y ∈ K ) ≤ Θ ^ + ε S D P 2 . ( ∵ Θ ^ ≥ Θ 1 ( S ) − ε S D P 4 ) \begin{aligned}
\sum_{j \in S} \frac{1+2Y_{rj}}{3} \le\Theta_1(S) + \frac{\varepsilon_{\mathrm{SDP}}}{4}
\qquad(\because Y \in K) \\
&\le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2}.
\qquad\left(\because\widehat{\Theta} \ge\Theta_1(S) - \frac{\varepsilon_{\mathrm{SDP}}}{4}\right)
\end{aligned} j ∈ S ∑ 3 1 + 2 Y r j ≤ Θ 1 ( S ) + 4 ε SDP ( ∵ Y ∈ K ) ≤ Θ + 2 ε SDP . ( ∵ Θ ≥ Θ 1 ( S ) − 4 ε SDP )
If m = 2 m = 2 m = 2 , then H = { Y ∈ X : ∑ j ∈ S 2 − 2 Y r j 3 ≤ Θ ^ + ε S D P 2 } \mathcal{H} = \{Y \in\mathcal{X} : \sum_{j \in S} \frac{2-2Y_{rj}}{3} \le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2}\} H = { Y ∈ X : ∑ j ∈ S 3 2 − 2 Y r j ≤ Θ + 2 ε SDP } , and every Y ∈ K Y \in K Y ∈ K satisfies
∑ j ∈ S 2 − 2 Y r j 3 ≤ Θ 2 ( S ) + ε S D P 4 ( ∵ Y ∈ K ) ≤ Θ ^ + ε S D P 2 . ( ∵ Θ ^ ≥ Θ 2 ( S ) − ε S D P 4 ) \begin{aligned}
\sum_{j \in S} \frac{2-2Y_{rj}}{3} \le\Theta_2(S) + \frac{\varepsilon_{\mathrm{SDP}}}{4}
\qquad(\because Y \in K) \\
&\le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2}.
\qquad\left(\because\widehat{\Theta} \ge\Theta_2(S) - \frac{\varepsilon_{\mathrm{SDP}}}{4}\right)
\end{aligned} j ∈ S ∑ 3 2 − 2 Y r j ≤ Θ 2 ( S ) + 4 ε SDP ( ∵ Y ∈ K ) ≤ Θ + 2 ε SDP . ( ∵ Θ ≥ Θ 2 ( S ) − 4 ε SDP )
Hence K ⊆ H K \subseteq\mathcal{H} K ⊆ H in both cases.
Next, suppose that O \mathcal{O} O returns ( X , S ) (X,\mathcal{S}) ( X , S ) , and let { v i } i ∈ V \{v_i\}_{i \in V} { v i } i ∈ V be a Gram realization of X X X . Since X ⪰ 0 X \succeq0 X ⪰ 0 , X i i = 1 X_{ii} = 1 X ii = 1 for every i ∈ V i \in V i ∈ V , and X X X satisfies the second to fourth constraints of K K K , the vectors { v i } i ∈ V \{v_i\}_{i \in V} { v i } i ∈ V are unit vectors satisfying the conditions on vectors in Theorem 34. Since no halfspace is returned by the check of the last family of constraints, we also have Γ r x ( X ) ⪰ − ε S D P I \Gamma_{rx}(X) \succeq-\varepsilon_{\mathrm{SDP}} I Γ r x ( X ) ⪰ − ε SDP I for every r x ∈ E rx \in E r x ∈ E . Fix r ∈ V r \in V r ∈ V and P ∈ P P \in\mathcal{P} P ∈ P , and let S = S r , P = { j ∈ V : ⟨ v r , v j ⟩ ∈ P } S = S_{r,P} = \{j \in V : \langle v_r, v_j\rangle\in P\} S = S r , P = { j ∈ V : ⟨ v r , v j ⟩ ∈ P } . Since no halfspace is returned for ( r , P ) (r,P) ( r , P ) , the witness W r , P \mathcal{W}_{r,P} W r , P with objective value Θ ^ \widehat{\Theta} Θ satisfies the following. If P P P is positive, then W r , P \mathcal{W}_{r,P} W r , P is feasible for Θ 1 ( S ) \Theta_1(S) Θ 1 ( S ) and ∑ j ∈ S 1 + 2 X r j 3 ≤ Θ ^ + ε S D P 2 \sum_{j \in S} \frac{1+2X_{rj}}{3} \le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2} ∑ j ∈ S 3 1 + 2 X r j ≤ Θ + 2 ε SDP . Since ρ P = inf P \rho_P = \inf P ρ P = inf P , we have ⟨ v r , v j ⟩ ≥ ρ P \langle v_r, v_j\rangle\ge\rho_P ⟨ v r , v j ⟩ ≥ ρ P for every j ∈ S j \in S j ∈ S , and hence
∑ j ∈ S ∥ x j ∥ 2 = Θ ^ ≥ ∑ j ∈ S 1 + 2 ⟨ v r , v j ⟩ 3 − ε S D P ≥ 1 + 2 ρ P 3 ∣ S ∣ − ε S D P , \begin{aligned}
\sum_{j \in S} \lVert x_j \rVert^2 = \widehat{\Theta} \\
&\ge\sum_{j \in S} \frac{1+2\langle v_r,v_j\rangle}{3} - \varepsilon_{\mathrm{SDP}} \\
&\ge\frac{1+2\rho_P}{3}|S| - \varepsilon_{\mathrm{SDP}},
\end{aligned} j ∈ S ∑ ∥ x j ∥ 2 = Θ ≥ j ∈ S ∑ 3 1 + 2 ⟨ v r , v j ⟩ − ε SDP ≥ 3 1 + 2 ρ P ∣ S ∣ − ε SDP ,
which is (i) of Theorem 34. If P P P is negative, then W r , P \mathcal{W}_{r,P} W r , P is feasible for Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) and ∑ j ∈ S 2 − 2 X r j 3 ≤ Θ ^ + ε S D P 2 \sum_{j \in S} \frac{2-2X_{rj}}{3} \le\widehat{\Theta} + \frac{\varepsilon_{\mathrm{SDP}}}{2} ∑ j ∈ S 3 2 − 2 X r j ≤ Θ + 2 ε SDP . Since ρ P = sup P \rho_P = \sup P ρ P = sup P , we have ⟨ v r , v j ⟩ ≤ ρ P \langle v_r, v_j\rangle\le\rho_P ⟨ v r , v j ⟩ ≤ ρ P for every j ∈ S j \in S j ∈ S , and hence
∑ j ∈ S ( ∥ x j , 1 ∥ 2 + ∥ x j , 2 ∥ 2 ) = Θ ^ ≥ ∑ j ∈ S 2 − 2 ⟨ v r , v j ⟩ 3 − ε S D P ≥ 2 − 2 ρ P 3 ∣ S ∣ − ε S D P , \begin{aligned}
\sum_{j \in S} \left(\lVert x_{j,1}\rVert^2 + \lVert x_{j,2}\rVert^2\right) = \widehat{\Theta} \\
&\ge\sum_{j \in S} \frac{2-2\langle v_r,v_j\rangle}{3} - \varepsilon_{\mathrm{SDP}} \\
&\ge\frac{2-2\rho_P}{3}|S| - \varepsilon_{\mathrm{SDP}},
\end{aligned} j ∈ S ∑ ( ∥ x j , 1 ∥ 2 + ∥ x j , 2 ∥ 2 ) = Θ ≥ j ∈ S ∑ 3 2 − 2 ⟨ v r , v j ⟩ − ε SDP ≥ 3 2 − 2 ρ P ∣ S ∣ − ε SDP ,
which is (ii) of Theorem 34.
Finally, the checks before the loop take polynomial time, and the loop calls SolveTheta at most ∣ V ∣ ∣ P ∣ = O ( n log n ) |V||\mathcal{P}| = O(n \log n) ∣ V ∣∣ P ∣ = O ( n log n ) times, each of which takes polynomial time. □
Proof of Theorem 34. First, we prove B ( X ~ , R 0 ) ⊆ K B(\widetilde{X},R_0) \subseteq K B ( X , R 0 ) ⊆ K . Write Y = X ~ + Δ X Y = \widetilde{X} + \Delta X Y = X + Δ X where Δ X \Delta X Δ X is an arbitrary perturbation with Δ X = Δ X T \Delta X = \Delta X^{\mathsf{T}} Δ X = Δ X T , Δ X i i = 0 \Delta X_{ii} = 0 Δ X ii = 0 and ∥ Δ X ∥ F ≤ R 0 \lVert\Delta X\rVert_F \le R_0 ∥ Δ X ∥ F ≤ R 0 . Thus, ∣ Δ X i j ∣ ≤ R 0 |\Delta X_{ij}| \le R_0 ∣Δ X ij ∣ ≤ R 0 and $\lVert \Delta X\rVert_{\mathrm{op}} \le ∥ Δ X ∥ F ≤ R 0 \|\Delta X\|_F \le R_0 ∥Δ X ∥ F ≤ R 0 hold. This gives Y = X ~ + Δ X ⪰ ( ε 0 − ε 0 8 n ) I ⪰ 0 Y=\widetilde{X}+\Delta X \succeq\left(\varepsilon_0-\frac{\varepsilon_0}{8n}\right)I \succeq0 Y = X + Δ X ⪰ ( ε 0 − 8 n ε 0 ) I ⪰ 0 . Since ( X ~ ) i j ≥ − 1 2 + ε 0 2 (\widetilde{X})_{ij}\ge-\frac{1}{2}+\frac{\varepsilon_0}{2} ( X ) ij ≥ − 2 1 + 2 ε 0 holds, we have Y i j ≥ − 1 2 + ε 0 2 − R 0 ≥ − 1 2 Y_{ij}\ge-\frac{1}{2}+\frac{\varepsilon_0}{2}-R_0\ge-\frac{1}{2} Y ij ≥ − 2 1 + 2 ε 0 − R 0 ≥ − 2 1 . Moreover, since ( X ~ ) i j = − 1 2 + ε 0 2 (\widetilde{X})_{ij}=-\frac{1}{2}+\frac{\varepsilon_0}{2} ( X ) ij = − 2 1 + 2 ε 0 for every edge i j ∈ E ij\in E ij ∈ E , we have Y i j ≤ − 1 2 + ε 0 2 + R 0 ≤ − 1 2 + ε S D P Y_{ij}\le-\frac{1}{2}+\frac{\varepsilon_0}{2}+R_0\le-\frac{1}{2}+\varepsilon_{\mathrm{SDP}} Y ij ≤ − 2 1 + 2 ε 0 + R 0 ≤ − 2 1 + ε SDP . For the fourth constraint, X i j + X i j ′ ≤ 1 2 X_{ij}+X_{ij'}\le\frac{1}{2} X ij + X i j ′ ≤ 2 1 holds and ( X ~ ) i j + ( X ~ ) i j ′ ≤ ( 1 − ε 0 ) 1 2 (\widetilde{X})_{ij}+(\widetilde{X})_{ij'}\le(1-\varepsilon_0)\frac{1}{2} ( X ) ij + ( X ) i j ′ ≤ ( 1 − ε 0 ) 2 1 . Therefore, Y i j + Y i j ′ ≤ ( 1 − ε 0 ) 1 2 + 2 R 0 ≤ 1 2 Y_{ij}+Y_{ij'}\le(1-\varepsilon_0)\frac{1}{2}+2R_0\le\frac{1}{2} Y ij + Y i j ′ ≤ ( 1 − ε 0 ) 2 1 + 2 R 0 ≤ 2 1 .
Since 1 + 2 X i j 3 \frac{1+2X_{ij}}{3} 3 1 + 2 X ij is 1 if j j j has the same color as i i i and 0 otherwise, ∑ j ∈ S 1 + 2 X i j 3 \sum_{j\in S}\frac{1+2X_{ij}}{3} ∑ j ∈ S 3 1 + 2 X ij is the number of vertices of S S S with the color of i i i , which form an independent set of G [ S ] G[S] G [ S ] . Similarly, ∑ j ∈ S 2 − 2 X i j 3 \sum_{j\in S}\frac{2-2X_{ij}}{3} ∑ j ∈ S 3 2 − 2 X ij is the number of the other vertices of S S S , which induce a 2-colorable subgraph of G [ S ] G[S] G [ S ] . Since Θ 1 ( S ) \Theta_1(S) Θ 1 ( S ) and Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) are relaxations of these problems, these sums are at most Θ 1 ( S ) \Theta_1(S) Θ 1 ( S ) and Θ 2 ( S ) \Theta_2(S) Θ 2 ( S ) , respectively. Hence, we obtain ∑ j ∈ S 1 + 2 ( X ~ ) i j 3 ≤ Θ 1 ( S ) + ε 0 n 3 \sum_{j\in S}\frac{1+2(\widetilde{X})_{ij}}{3}\le\Theta_1(S)+\frac{\varepsilon_0n}{3} ∑ j ∈ S 3 1 + 2 ( X ) ij ≤ Θ 1 ( S ) + 3 ε 0 n and ∑ j ∈ S 2 − 2 ( X ~ ) i j 3 ≤ Θ 2 ( S ) + 2 ε 0 n 3 \sum_{j\in S}\frac{2-2(\widetilde{X})_{ij}}{3}\le\Theta_2(S)+\frac{2\varepsilon_0n}{3} ∑ j ∈ S 3 2 − 2 ( X ) ij ≤ Θ 2 ( S ) + 3 2 ε 0 n . Moreover,
∣ ∑ j ∈ S 1 + 2 Y i j 3 − ∑ j ∈ S 1 + 2 ( X ~ ) i j 3 ∣ ≤ 2 ∣ S ∣ R 0 3 ≤ ε 0 12 \left|\sum_{j\in S}\frac{1+2Y_{ij}}{3}-\sum_{j\in S}\frac{1+2(\widetilde{X})_{ij}}{3}\right|\le\frac{2|S|R_0}{3}\le\frac{\varepsilon_0}{12} j ∈ S ∑ 3 1 + 2 Y ij − j ∈ S ∑ 3 1 + 2 ( X ) ij ≤ 3 2∣ S ∣ R 0 ≤ 12 ε 0
and
∣ ∑ j ∈ S 2 − 2 Y i j 3 − ∑ j ∈ S 2 − 2 ( X ~ ) i j 3 ∣ ≤ 2 ∣ S ∣ R 0 3 ≤ ε 0 12 . \left|\sum_{j\in S}\frac{2-2Y_{ij}}{3}-\sum_{j\in S}\frac{2-2(\widetilde{X})_{ij}}{3}\right|\le\frac{2|S|R_0}{3}\le\frac{\varepsilon_0}{12}. j ∈ S ∑ 3 2 − 2 Y ij − j ∈ S ∑ 3 2 − 2 ( X ) ij ≤ 3 2∣ S ∣ R 0 ≤ 12 ε 0 .
Thus, ∑ j ∈ S 1 + 2 Y i j 3 ≤ Θ 1 ( S ) + ε 0 n ≤ Θ 1 ( S ) + ε S D P 4 \sum_{j\in S}\frac{1+2Y_{ij}}{3}\le\Theta_1(S)+\varepsilon_0n\le\Theta_1(S)+\frac{\varepsilon_{\mathrm{SDP}}}{4} ∑ j ∈ S 3 1 + 2 Y ij ≤ Θ 1 ( S ) + ε 0 n ≤ Θ 1 ( S ) + 4 ε SDP and ∑ j ∈ S 2 − 2 Y i j 3 ≤ Θ 2 ( S ) + ε 0 n ≤ Θ 2 ( S ) + ε S D P 4 \sum_{j\in S}\frac{2-2Y_{ij}}{3}\le\Theta_2(S)+\varepsilon_0n\le\Theta_2(S)+\frac{\varepsilon_{\mathrm{SDP}}}{4} ∑ j ∈ S 3 2 − 2 Y ij ≤ Θ 2 ( S ) + ε 0 n ≤ Θ 2 ( S ) + 4 ε SDP .
It remains to check the last family of constraints. The matrix X ‾ \overline{X} X is the Gram matrix of { e c ( i ) } i ∈ V \{e_{c(i)}\}_{i\in V} { e c ( i ) } i ∈ V , where c ( i ) c(i) c ( i ) is the color of i i i and e 1 , e 2 , e 3 ∈ R 2 e_1,e_2,e_3\in\mathbb{R}^2 e 1 , e 2 , e 3 ∈ R 2 are unit vectors with pairwise inner products − 1 2 -\frac{1}{2} − 2 1 . Fix an edge r x ∈ E rx\in E r x ∈ E . Let d d d be the color other than c ( r ) c(r) c ( r ) and c ( x ) c(x) c ( x ) . For y ∈ N V ( r ) y\in N_V(r) y ∈ N V ( r ) , we have c ( y ) ∈ { c ( x ) , d } c(y)\in\{c(x),d\} c ( y ) ∈ { c ( x ) , d } , and since e 1 + e 2 + e 3 = 0 e_1+e_2+e_3=0 e 1 + e 2 + e 3 = 0 , we get ( 2 e c ( y ) + e c ( r ) ) / 3 = ± ( e c ( x ) − e d ) / 3 (2e_{c(y)}+e_{c(r)})/\sqrt{3}=\pm(e_{c(x)}-e_d)/\sqrt{3} ( 2 e c ( y ) + e c ( r ) ) / 3 = ± ( e c ( x ) − e d ) / 3 ; similarly, ( 2 e c ( z ) + e c ( x ) ) / 3 = ± ( e c ( r ) − e d ) / 3 (2e_{c(z)}+e_{c(x)})/\sqrt{3}=\pm(e_{c(r)}-e_d)/\sqrt{3} ( 2 e c ( z ) + e c ( x ) ) / 3 = ± ( e c ( r ) − e d ) / 3 for z ∈ N V ( x ) z\in N_V(x) z ∈ N V ( x ) . Thus, for α 1 ∈ R N V ( r ) \alpha_1\in\mathbb{R}^{N_V(r)} α 1 ∈ R N V ( r ) and α 2 ∈ R N V ( x ) \alpha_2\in\mathbb{R}^{N_V(x)} α 2 ∈ R N V ( x ) , the vectors u 1 = ∑ y ( α 1 ) y ( 2 e c ( y ) + e c ( r ) ) / 3 u_1=\sum_y(\alpha_1)_y(2e_{c(y)}+e_{c(r)})/\sqrt{3} u 1 = ∑ y ( α 1 ) y ( 2 e c ( y ) + e c ( r ) ) / 3 and u 2 = ∑ z ( α 2 ) z ( 2 e c ( z ) + e c ( x ) ) / 3 u_2=\sum_z(\alpha_2)_z(2e_{c(z)}+e_{c(x)})/\sqrt{3} u 2 = ∑ z ( α 2 ) z ( 2 e c ( z ) + e c ( x ) ) / 3 lie in R ( e c ( x ) − e d ) \mathbb{R}(e_{c(x)}-e_d) R ( e c ( x ) − e d ) and R ( e c ( r ) − e d ) \mathbb{R}(e_{c(r)}-e_d) R ( e c ( r ) − e d ) , respectively. Since ∥ e c ( x ) − e d ∥ 2 = ∥ e c ( r ) − e d ∥ 2 = 3 \|e_{c(x)}-e_d\|^2=\|e_{c(r)}-e_d\|^2=3 ∥ e c ( x ) − e d ∥ 2 = ∥ e c ( r ) − e d ∥ 2 = 3 and ⟨ e c ( x ) − e d , e c ( r ) − e d ⟩ = 3 2 \langle e_{c(x)}-e_d,e_{c(r)}-e_d\rangle=\frac{3}{2} ⟨ e c ( x ) − e d , e c ( r ) − e d ⟩ = 2 3 , we have ∣ ⟨ u 1 , u 2 ⟩ ∣ = 1 2 ∥ u 1 ∥ ∥ u 2 ∥ |\langle u_1,u_2\rangle|=\frac{1}{2}\|u_1\|\|u_2\| ∣ ⟨ u 1 , u 2 ⟩ ∣ = 2 1 ∥ u 1 ∥∥ u 2 ∥ , and hence
( α 1 , α 2 ) T Γ r x ( X ‾ ) ( α 1 , α 2 ) = ∥ u 1 ∥ 2 + 4 ⟨ u 1 , u 2 ⟩ + ∥ u 2 ∥ 2 ≥ ( ∥ u 1 ∥ − ∥ u 2 ∥ ) 2 ≥ 0. (\alpha_1,\alpha_2)^\mathsf{T}\Gamma_{rx}(\overline{X})(\alpha_1,\alpha_2)=\|u_1\|^2+4\langle u_1,u_2\rangle+\|u_2\|^2\ge(\|u_1\|-\|u_2\|)^2\ge0. ( α 1 , α 2 ) T Γ r x ( X ) ( α 1 , α 2 ) = ∥ u 1 ∥ 2 + 4 ⟨ u 1 , u 2 ⟩ + ∥ u 2 ∥ 2 ≥ ( ∥ u 1 ∥ − ∥ u 2 ∥ ) 2 ≥ 0.
Thus Γ r x ( X ‾ ) ⪰ 0 \Gamma_{rx}(\overline{X})\succeq0 Γ r x ( X ) ⪰ 0 . Recall that, for Z ∈ X Z\in\mathcal{X} Z ∈ X , the entry of Γ r x ( Z ) \Gamma_{rx}(Z) Γ r x ( Z ) indexed by ( ( i , j ) , ( i ′ , j ′ ) ) ((i,j),(i',j')) (( i , j ) , ( i ′ , j ′ )) is 1 + 1 i ≠ i ′ 3 ( 4 Z j j ′ + 2 Z j i ′ + 2 Z i j ′ + Z i i ′ ) \frac{1+1_{i\ne i'}}{3}(4Z_{jj'}+2Z_{ji'}+2Z_{ij'}+Z_{ii'}) 3 1 + 1 i = i ′ ( 4 Z j j ′ + 2 Z j i ′ + 2 Z i j ′ + Z i i ′ ) . For every Z ∈ X Z\in\mathcal{X} Z ∈ X with ∣ Z i j ∣ ≤ 1 |Z_{ij}|\le1 ∣ Z ij ∣ ≤ 1 , every entry of Γ r x ( Z ) \Gamma_{rx}(Z) Γ r x ( Z ) has absolute value at most 2 3 ( 4 + 2 + 2 + 1 ) = 6 \frac{2}{3}(4+2+2+1)=6 3 2 ( 4 + 2 + 2 + 1 ) = 6 , and Γ r x ( Z ) \Gamma_{rx}(Z) Γ r x ( Z ) has at most 2 n 2n 2 n rows, so ∥ Γ r x ( I ) ∥ o p ≤ ∥ Γ r x ( I ) ∥ F ≤ 12 n \|\Gamma_{rx}(I)\|_{\mathrm{op}}\le\|\Gamma_{rx}(I)\|_F\le12n ∥ Γ r x ( I ) ∥ op ≤ ∥ Γ r x ( I ) ∥ F ≤ 12 n . Since Γ r x \Gamma_{rx} Γ r x is affine, Γ r x ( X ~ ) = ( 1 − ε 0 ) Γ r x ( X ‾ ) + ε 0 Γ r x ( I ) ⪰ − 12 n ε 0 I = − 3 4 ε S D P I \Gamma_{rx}(\widetilde{X})=(1-\varepsilon_0)\Gamma_{rx}(\overline{X})+\varepsilon_0\Gamma_{rx}(I)\succeq-12n\varepsilon_0I=-\frac{3}{4}\varepsilon_{\mathrm{SDP}}I Γ r x ( X ) = ( 1 − ε 0 ) Γ r x ( X ) + ε 0 Γ r x ( I ) ⪰ − 12 n ε 0 I = − 4 3 ε SDP I . Moreover, every entry of Γ r x ( Y ) − Γ r x ( X ~ ) \Gamma_{rx}(Y)-\Gamma_{rx}(\widetilde{X}) Γ r x ( Y ) − Γ r x ( X ) has absolute value at most 2 3 ( 4 + 2 + 2 + 1 ) R 0 = 6 R 0 \frac{2}{3}(4+2+2+1)R_0=6R_0 3 2 ( 4 + 2 + 2 + 1 ) R 0 = 6 R 0 , so ∥ Γ r x ( Y ) − Γ r x ( X ~ ) ∥ o p ≤ 12 n R 0 = 3 2 ε 0 ≤ 1 4 ε S D P \|\Gamma_{rx}(Y)-\Gamma_{rx}(\widetilde{X})\|_{\mathrm{op}}\le12nR_0=\frac{3}{2}\varepsilon_0\le\frac{1}{4}\varepsilon_{\mathrm{SDP}} ∥ Γ r x ( Y ) − Γ r x ( X ) ∥ op ≤ 12 n R 0 = 2 3 ε 0 ≤ 4 1 ε SDP . Hence Γ r x ( Y ) ⪰ − ε S D P I \Gamma_{rx}(Y)\succeq-\varepsilon_{\mathrm{SDP}}I Γ r x ( Y ) ⪰ − ε SDP I . Therefore, we obtain B ( X ~ , R 0 ) ⊆ K B(\widetilde{X},R_0)\subseteq K B ( X , R 0 ) ⊆ K .
Next, we prove K ⊆ B ( I , R 1 ) K\subseteq B(I,R_1) K ⊆ B ( I , R 1 ) . Let X ∈ K X\in K X ∈ K . Since X ⪰ 0 X\succeq0 X ⪰ 0 and X i i = 1 X_{ii}=1 X ii = 1 for every i ∈ V i\in V i ∈ V , we have ∣ X i j ∣ ≤ 1 |X_{ij}|\le1 ∣ X ij ∣ ≤ 1 for all i ≠ j i\ne j i = j . Thus, ∥ X − I ∥ F 2 = 2 ∑ i < j X i j 2 ≤ 2 ( n 2 ) \|X-I\|_F^2=2\sum_{i<j}X_{ij}^2\le2\binom{n}{2} ∥ X − I ∥ F 2 = 2 ∑ i < j X ij 2 ≤ 2 ( 2 n ) , which gives ∥ X − I ∥ F ≤ 2 ( n 2 ) ≤ n \|X-I\|_F\le\sqrt{2\binom{n}{2}}\le n ∥ X − I ∥ F ≤ 2 ( 2 n ) ≤ n . This implies that K ⊆ B ( I , R 1 ) K\subseteq B(I,R_1) K ⊆ B ( I , R 1 ) .
By the standard analysis of the ellipsoid method [27 ] , in Algorithm 6, the OUTPUTORCUT oracle is called at most O ( n 4 log ( R 1 / R 0 ) ) = O ( n 4 log ( n / ε S D P ) ) O(n^4\log(R_1/R_0))=O(n^4\log(n/\varepsilon_{\mathrm{SDP}})) O ( n 4 log ( R 1 / R 0 )) = O ( n 4 log ( n / ε SDP )) times, since X \mathcal{X} X has dimension n ( n − 1 ) / 2 n(n-1)/2 n ( n − 1 ) /2 . Hence, by Theorem 59, Algorithm 6 returns an output satisfying Theorem 34. □ \square □
Justification of the Exact-Constraint Assumption
First, we justify treating the strict vector coloring constraints as exact. The computed SDP solution satisfies these constraints up to ε S D P \varepsilon_{\mathrm{SDP}} ε SDP . This induces O ( ε S D P ) O(\varepsilon_{\mathrm{SDP}}) O ( ε SDP ) perturbations in v x y v_{xy} v x y and w y w_y w y , and hence O ( n ε S D P ) O(n\varepsilon_{\mathrm{SDP}}) O ( n ε SDP ) perturbations in their Gaussian projections with probability at least 1 − e − n 1-e^{-n} 1 − e − n . These errors are absorbed either by the slack already present in the thresholds or into the o ( t 2 ) o(t^2) o ( t 2 ) terms in the exponents of the probability bounds throughout our analysis.
Next, we discuss the semidefinite constraint. In the proof of Theorem 38, we treat the approximate constraint Γ r x ( X ) ⪰ − ε S D P I \Gamma_{rx}(X) \succeq-\varepsilon_{\mathrm{SDP}}I Γ r x ( X ) ⪰ − ε SDP I as if Γ r x ( X ) ⪰ 0 \Gamma_{rx}(X) \succeq0 Γ r x ( X ) ⪰ 0 . This requires justification even when ε S D P \varepsilon_{\mathrm{SDP}} ε SDP is arbitrarily small. Indeed, if the vectors are nearly linearly dependent, an additive error in the semidefinite constraint need not imply a small error in the correlation bound of Theorem 41 .
We address this issue by adding tiny independent Gaussian noise to the Gaussian projections used in the proof of (ii) of Theorem 42 . Fix a vertex r r r . For every y ∈ N V ( r ) y \in N_V(r) y ∈ N V ( r ) , and every x ∈ N V ( r ) x \in N_V(r) x ∈ N V ( r ) and z ∈ N V ( x ) z \in N_V(x) z ∈ N V ( x ) , let h r y h_{ry} h r y and h x z h_{xz} h x z be mutually independent standard Gaussian variables, also independent of g g g . Define
T r y = ⟨ g , v r y ⟩ + σ h r y , T_{ry} = \langle g, v_{ry}\rangle+ \sigma h_{ry}, T r y = ⟨ g , v r y ⟩ + σ h r y ,
T x z = ⟨ g , v x z ⟩ + σ h x z . T_{xz} = \langle g, v_{xz}\rangle+ \sigma h_{xz}. T x z = ⟨ g , v x z ⟩ + σ h x z .
where σ = e − n \sigma=e^{-n} σ = e − n . Using v x z = ξ z v x r + w z v_{xz}=\xi_zv_{xr}+w_z v x z = ξ z v x r + w z , we further define
Z x z = T x z − ξ z T x r = ⟨ g , w z ⟩ + σ ( h x z − ξ z h x r ) . Z_{xz}=T_{xz}-\xi_zT_{xr}=\langle g,w_z\rangle+\sigma(h_{xz}-\xi_zh_{xr}). Z x z = T x z − ξ z T x r = ⟨ g , w z ⟩ + σ ( h x z − ξ z h x r ) .
Moreover,
E [ T r x Z x z ] = ⟨ v r x , w z ⟩ = 0 \mathbb{E}[T_{rx}Z_{xz}]=\langle v_{rx},w_z\rangle=0 E [ T r x Z x z ] = ⟨ v r x , w z ⟩ = 0
holds for every z ∈ N V ( x ) z\in N_V(x) z ∈ N V ( x ) . Since each edge residual v i j v_{ij} v ij (i j ∈ E ij\in E ij ∈ E ) differs from ( 2 v j + v i ) / 3 (2v_j+v_i)/\sqrt{3} ( 2 v j + v i ) / 3 by O ( ε S D P ) O(\varepsilon_{\mathrm{SDP}}) O ( ε SDP ) , the matrix
Γ ~ r x = ( E [ T r T r T ] 2 E [ T r T x T ] 2 E [ T x T r T ] E [ T x T x T ] ) \widetilde{\Gamma}_{rx} =
\begin{pmatrix}
\mathbb{E}[T_rT_r^{\mathsf T}] & 2\mathbb{E}[T_rT_x^{\mathsf T}] \\
2\mathbb{E}[T_xT_r^{\mathsf T}] & \mathbb{E}[T_xT_x^{\mathsf T}]
\end{pmatrix} Γ r x = ( E [ T r T r T ] 2 E [ T x T r T ] 2 E [ T r T x T ] E [ T x T x T ] )
differs from Γ r x ( X ) + σ 2 I \Gamma_{rx}(X)+\sigma^2I Γ r x ( X ) + σ 2 I by O ( n ε S D P ) O(n\varepsilon_{\mathrm{SDP}}) O ( n ε SDP ) in operator norm, where T r = ( T r y ) y ∈ N V ( r ) T_r=(T_{ry})_{y\in N_V(r)} T r = ( T r y ) y ∈ N V ( r ) and T x = ( T x z ) z ∈ N V ( x ) T_x=(T_{xz})_{z\in N_V(x)} T x = ( T x z ) z ∈ N V ( x ) . Since Γ r x ( X ) ⪰ − ε S D P I \Gamma_{rx}(X)\succeq-\varepsilon_{\mathrm{SDP}}I Γ r x ( X ) ⪰ − ε SDP I and n ε S D P = n e − 3 n = o ( σ 2 ) n\varepsilon_{\mathrm{SDP}}=ne^{-3n}=o(\sigma^2) n ε SDP = n e − 3 n = o ( σ 2 ) , we have Γ ~ r x ⪰ 0 \widetilde{\Gamma}_{rx}\succeq0 Γ r x ⪰ 0 for sufficiently large n n n . Hence the argument of Theorem 41 gives ∣ E [ u v ] ∣ ≤ 1 2 E [ u 2 ] E [ v 2 ] |\mathbb{E}[uv]|\leq\frac{1}{2}\sqrt{\mathbb{E}[u^2]\mathbb{E}[v^2]} ∣ E [ uv ] ∣ ≤ 2 1 E [ u 2 ] E [ v 2 ] for all u ∈ span { T r y : y ∈ N V ( r ) } u\in\operatorname{span}\{T_{ry}:y\in N_V(r)\} u ∈ span { T r y : y ∈ N V ( r )} and v ∈ span { T x z : z ∈ N V ( x ) } v\in\operatorname{span}\{T_{xz}:z\in N_V(x)\} v ∈ span { T x z : z ∈ N V ( x )} .
Since E [ T r x Z x z ] = 0 \mathbb{E}[T_{rx}Z_{xz}]=0 E [ T r x Z x z ] = 0 for every z ∈ N V ( x ) z\in N_V(x) z ∈ N V ( x ) and these variables are jointly Gaussian, the family { Z x z } z ∈ N V ( x ) \{Z_{xz}\}_{z\in N_V(x)} { Z x z } z ∈ N V ( x ) is independent of T r x T_{rx} T r x . Moreover, writing
Z r y = T r y − E [ T r y T r x ] E [ T r x 2 ] T r x Z_{ry}=T_{ry}-\frac{\mathbb{E}[T_{ry}T_{rx}]}{\mathbb{E}[T_{rx}^2]}T_{rx} Z r y = T r y − E [ T r x 2 ] E [ T r y T r x ] T r x
for y ∈ N V ( r ) y\in N_V(r) y ∈ N V ( r ) , the family { Z r y } y ∈ N V ( r ) \{Z_{ry}\}_{y\in N_V(r)} { Z r y } y ∈ N V ( r ) is also independent of T r x T_{rx} T r x . Since Z r y ∈ span { T r y ′ : y ′ ∈ N V ( r ) } Z_{ry}\in\operatorname{span}\{T_{ry'}:y'\in N_V(r)\} Z r y ∈ span { T r y ′ : y ′ ∈ N V ( r )} and Z x z ∈ span { T x z ′ : z ′ ∈ N V ( x ) } Z_{xz}\in\operatorname{span}\{T_{xz'}:z'\in N_V(x)\} Z x z ∈ span { T x z ′ : z ′ ∈ N V ( x )} (recall that r ∈ N V ( x ) r\in N_V(x) r ∈ N V ( x ) ), the correlation bound above remains valid between the spans of { Z r y } y \{Z_{ry}\}_y { Z r y } y and { Z x z } z \{Z_{xz}\}_z { Z x z } z . These spans play the roles of the subspaces U r ∩ span ( v r , v x ) ⊥ \mathcal{U}_r\cap\operatorname{span}(v_r,v_x)^\perp U r ∩ span ( v r , v x ) ⊥ and U x ∩ span ( v r , v x ) ⊥ \mathcal{U}_x\cap\operatorname{span}(v_r,v_x)^\perp U x ∩ span ( v r , v x ) ⊥ in the proof of (ii) of Theorem 42 .
Accordingly, we define the maximizer events and the events B ~ x − \widetilde{B}_x^{-} B x − by the noisy variables. Put L ~ = { max y ∈ N F j ( r ) T r y > A t } \widetilde{\mathcal{L}}=\{\max_{y\in N_{F_j}(r)}T_{ry}>At\} L = { max y ∈ N F j ( r ) T r y > A t } , let L ~ ( x ) \widetilde{\mathcal{L}}^{(x)} L ( x ) be the event that L ~ \widetilde{\mathcal{L}} L occurs and x x x is the maximizer of T r y T_{ry} T r y , and, in the setting of the induction step for the local bound in the proof of Theorem 38 (so that j = i + 2 j=i+2 j = i + 2 , A = A i + 1 A=A_{i+1} A = A i + 1 and A ′ = A i A'=A_i A ′ = A i ), put
B ~ x − = { ∀ z ∈ N F j − 1 ( x ) : − Z x z ≤ ( 1 + θ ) ( 3 a − b ξ z ) t − θ ( A ′ − c ∣ ξ z ∣ ) t − 1 + θ t } . \widetilde{\mathcal{B}}_x^-=\left\{\forall z\in N_{F_{j-1}}(x):-Z_{xz}\leq(1+\theta)(\sqrt{3}a-b\xi_z)t-\theta(A'-c|\xi_z|)t-\frac{1+\theta}{t}\right\}. B x − = { ∀ z ∈ N F j − 1 ( x ) : − Z x z ≤ ( 1 + θ ) ( 3 a − b ξ z ) t − θ ( A ′ − c ∣ ξ z ∣ ) t − t 1 + θ } .
which is the event B x − \mathcal{B}_x^- B x − in the proof of Theorem 42 with − ⟨ g ⊥ , w z ⟩ -\langle g_\perp,w_z\rangle − ⟨ g ⊥ , w z ⟩ replaced by − Z x z -Z_{xz} − Z x z . With probability at least 1 − e − Ω ( n 2 ) 1-e^{-\Omega(n^2)} 1 − e − Ω ( n 2 ) , all the introduced Gaussian noises have absolute value at most n n n . On this event, the perturbations in the Gaussian projections are O ( n σ ) O(n\sigma) O ( nσ ) . These perturbations are absorbed by the existing slack in the measure estimates and inequalities, and the exceptional probability is negligible compared with ε t \varepsilon_t ε t . Since T r y = E [ T r y T r x ] E [ T r x 2 ] T r x + Z r y T_{r y}=\frac{\mathbb{E}[T_{r y}T_{r x}]}{\mathbb{E}[T_{r x}^2]}T_{r x}+Z_{r y} T r y = E [ T r x 2 ] E [ T r y T r x ] T r x + Z r y , conditional on T r x = s T_{r x}=s T r x = s , the event L ~ ( x ) \widetilde{\mathcal{L}}^{(x)} L ( x ) depends only on { Z r y } y ∈ N V ( r ) \{Z_{r y}\}_{y\in N_V(r)} { Z r y } y ∈ N V ( r ) , while B ~ x − \widetilde{\mathcal{B}}_x^- B x − depends only on { Z x z } z ∈ N V ( x ) \{Z_{x z}\}_{z\in N_V(x)} { Z x z } z ∈ N V ( x ) and is independent of T r x T_{r x} T r x . Hence, conditional on T r x T_{r x} T r x , Theorem 16 applies to L ~ ( x ) \widetilde{\mathcal{L}}^{(x)} L ( x ) and B ~ x − \widetilde{\mathcal{B}}_x^- B x − with the same exponents p 1 , p 2 p_1,p_2 p 1 , p 2 as in the proof of Theorem 38, since only the correlation bound 1 2 \frac{1}{2} 2 1 is used, and the proof of (ii) of Theorem 42 goes through with L ~ \widetilde{\mathcal{L}} L , L ~ ( x ) \widetilde{\mathcal{L}}^{(x)} L ( x ) and B ~ x − \widetilde{\mathcal{B}}_x^- B x − in place of L \mathcal{L} L , L ( x ) \mathcal{L}^{(x)} L ( x ) and B x − \mathcal{B}_x^- B x − , and with P [ T r x > A t ] = Φ ‾ ( A t / 1 + σ 2 ) ≤ 2 Φ ‾ ( A t ) \mathbb{P}[T_{r x}>At]=\overline{\Phi}(At/\sqrt{1+\sigma^2})\leq2\overline{\Phi}(At) P [ T r x > A t ] = Φ ( A t / 1 + σ 2 ) ≤ 2 Φ ( A t ) in place of Φ ‾ ( A t ) \overline{\Phi}(At) Φ ( A t ) . In the application of (ii) in the proof of Theorem 38, the resulting extra factor 2 1 − 1 / p 1 = 2 1 / 11997 2^{1-1/p_1}=2^{1/11997} 2 1 − 1/ p 1 = 2 1/11997 is absorbed by the factor e − η t 2 / 11997 e^{-\eta t^2/11997} e − η t 2 /11997 there for sufficiently large t t t .
Finally, we transfer the bound on the perturbed event L ~ \widetilde{\mathcal{L}} L to the original local event L \mathcal{L} L . Since
P [ L △ L ~ ] ≤ ∑ x ∈ N F j ( r ) P [ ∣ ⟨ g , v r x ⟩ − A t ∣ ≤ σ ∣ h r x ∣ ] ≤ ∑ x ∈ N F j ( r ) E h r x [ 2 σ ∣ h r x ∣ 2 π ] ≤ n 2 π 2 σ 2 π = O ( n e − n ) , \mathbb{P}[\mathcal{L}\triangle\widetilde{\mathcal{L}}]\leq\sum_{x\in N_{F_j}(r)}\mathbb{P}\left[\left|\langle g,v_{rx}\rangle-At\right|\leq\sigma|h_{rx}|\right]\leq\sum_{x\in N_{F_j}(r)}\mathbb{E}_{h_{rx}}\left[\frac{2\sigma|h_{rx}|}{\sqrt{2\pi}}\right]\leq n\sqrt{\frac{2}{\pi}}\frac{2\sigma}{\sqrt{2\pi}}=O(ne^{-n}), P [ L △ L ] ≤ x ∈ N F j ( r ) ∑ P [ ∣ ⟨ g , v r x ⟩ − A t ∣ ≤ σ ∣ h r x ∣ ] ≤ x ∈ N F j ( r ) ∑ E h r x [ 2 π 2 σ ∣ h r x ∣ ] ≤ n π 2 2 π 2 σ = O ( n e − n ) ,
we have P [ L ] ≤ P [ L ~ ] + O ( n e − n ) \mathbb{P}[\mathcal{L}]\leq\mathbb{P}[\widetilde{\mathcal{L}}]+O(ne^{-n}) P [ L ] ≤ P [ L ] + O ( n e − n ) and the same induction argument applies in Theorem 38.
Computational Verification
In this appendix, we prove Theorem 40, recalled as follows, and the property (P2) of Subsection 5.4, which states that if A = A 0 A=A_0 A = A 0 , then q c ≤ f A q_c\leq f_A q c ≤ f A holds, that is, q A = q c q_A=q_c q A = q c .
Lemma 60. Let 1.968 ≤ A ≤ A 0 1.968\leq A\leq A_0 1.968 ≤ A ≤ A 0 , and let ρ ∈ [ − 1 2 , 1 ) \rho\in[-\frac{1}{2},1) ρ ∈ [ − 2 1 , 1 ) satisfy the assumption on ( A , ρ ) (A,\rho) ( A , ρ ) in one of the two cases in the table below. Then there exist ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) , independent of ρ ′ \rho' ρ ′ , such that
(i) a , μ , c , κ ≥ 0 a,\mu,c,\kappa\geq0 a , μ , c , κ ≥ 0 , b ∈ R b\in\mathbb{R} b ∈ R and 0 < θ < 14 0<\theta<14 0 < θ < 14 ;
(ii) a 2 + b 2 + μ 2 ≤ 1 − 10 − 12 a^2+b^2+\mu^2\leq1-10^{-12} a 2 + b 2 + μ 2 ≤ 1 − 1 0 − 12 and c 2 + κ 2 ≤ 1 4 − 10 − 20 c^2+\kappa^2\leq\frac{1}{4}-10^{-20} c 2 + κ 2 ≤ 4 1 − 1 0 − 20 ;
(iii) L 2 + Λ ( μ + θ ( μ + κ ) ) 2 ≥ J + 10 − 18 L^2+\Lambda(\mu+\theta(\mu+\kappa))^2\geq J+10^{-18} L 2 + Λ ( μ + θ ( μ + κ ) ) 2 ≥ J + 1 0 − 18 ;
(iv) for every ρ ′ \rho' ρ ′ satisfying − 1 2 ≤ ρ ′ ≤ 1 2 − ρ -\frac{1}{2}\leq\rho'\leq\frac{1}{2}-\rho − 2 1 ≤ ρ ′ ≤ 2 1 − ρ and 3 4 − ρ 2 − ( ρ ′ ) 2 − ρ ρ ′ ≥ 0 \frac{3}{4}-\rho^2-(\rho')^2-\rho\rho'\geq0 4 3 − ρ 2 − ( ρ ′ ) 2 − ρ ρ ′ ≥ 0 ,
L ( 1 / 2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ / 2 ) + θ c ∣ ρ ′ + ρ / 2 ∣ 1 − ρ 2 + 3 2 ( 1 + θ ) a − 3 2 θ A ≥ q A ( ρ ′ ) + 3 ⋅ 10 − 6 . \frac{L(1/2+\rho\rho')-(1+\theta)b(\rho'+\rho/2)+\theta c|\rho'+\rho/2|}{\sqrt{1-\rho^2}}+\frac{3}{2}(1+\theta)a-\frac{\sqrt{3}}{2}\theta A\geq q_A(\rho')+3\cdot10^{-6}. 1 − ρ 2 L ( 1/2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ /2 ) + θ c ∣ ρ ′ + ρ /2∣ + 2 3 ( 1 + θ ) a − 2 3 θ A ≥ q A ( ρ ′ ) + 3 ⋅ 1 0 − 6 . :::`
Here ( L , J , Λ ) (L,J,\Lambda) ( L , J , Λ ) are as follows:
Table 1 (PDF p. 74).
First, we restrict the domain of Case 1; as illustrated in Figure 6 , not all ( A , ρ ) (A,\rho) ( A , ρ ) meet the assumption of Case 1. The assumption of Case 1 is q A ( ρ ) < q c ( ρ ) q_A(\rho)<q_c(\rho) q A ( ρ ) < q c ( ρ ) , so it is not required to verify the property for the range of ( A , ρ ) (A,\rho) ( A , ρ ) such that q A ( ρ ) = q c ( ρ ) q_A(\rho)=q_c(\rho) q A ( ρ ) = q c ( ρ ) , that is, f A ( ρ ) ≥ q c ( ρ ) f_A(\rho)\geq q_c(\rho) f A ( ρ ) ≥ q c ( ρ ) . With the following discussion, we also prove the property (P2).
Figure 6. The graph showing the value of f A ( ρ ) − q c ( ρ ) f_A(\rho)-q_c(\rho) f A ( ρ ) − q c ( ρ ) over ( A , ρ ) ∈ [ 1.968 , A 0 ] × [ − 1 2 , 1 ] (A,\rho)\in[1.968,A_0]\times[-\frac{1}{2},1] ( A , ρ ) ∈ [ 1.968 , A 0 ] × [ − 2 1 , 1 ] . The blue range satisfies the assumption of Case 1.
Lemma 61. Suppose 1.968 ≤ A ≤ A 0 1.968\leq A\leq A_0 1.968 ≤ A ≤ A 0 . If ρ ≥ 0.7 \rho\geq0.7 ρ ≥ 0.7 , then q A ( ρ ) = q c ( ρ ) q_A(\rho)=q_c(\rho) q A ( ρ ) = q c ( ρ ) , that is, f A ( ρ ) ≥ q c ( ρ ) f_A(\rho)\geq q_c(\rho) f A ( ρ ) ≥ q c ( ρ ) .
Proof. Assume ρ > 1 4 \rho>\frac{1}{4} ρ > 4 1 . We define P ( ρ ) : = f A ( ρ ) − f ~ ( A ) 1 − ρ 2 = − 0.8 ζ 2 + 0.45 ζ 2 2 + 0.39 ζ 3 + 0.65 ζ 2 3 P(\rho):=f_A(\rho)-\tilde{f}(A)\sqrt{1-\rho^2}=-0.8\zeta_2+0.45\zeta_2^2+0.39\zeta_3+0.65\zeta_2^3 P ( ρ ) := f A ( ρ ) − f ~ ( A ) 1 − ρ 2 = − 0.8 ζ 2 + 0.45 ζ 2 2 + 0.39 ζ 3 + 0.65 ζ 2 3 . Then, the following holds:
f A ( ρ ) − q c ( ρ ) = ( f ~ ( A ) 1 − ρ 2 + P ( ρ ) ) − ( 1 − ρ 2 ) ( C D a ( ρ ) + 1 4 ) = ( f ~ ( A ) − D a + 5 4 ) 1 − ρ 2 + P ( ρ ) ( ∵ C D a ( ρ ) = D a + 1 ) \begin{aligned}
f_A(\rho)-q_c(\rho)=\left(\tilde{f}(A)\sqrt{1-\rho^2}+P(\rho)\right)-\sqrt{(1-\rho^2)\left(CD_a(\rho)+\frac{1}{4}\right)} \\
&=\left(\widetilde{f}(A)-\sqrt{D_a+\frac{5}{4}}\right)\sqrt{1-\rho^2}+P(\rho)
\qquad\left(\because C_{D_a}(\rho)=D_a+1\right)
\end{aligned} f A ( ρ ) − q c ( ρ ) = ( f ~ ( A ) 1 − ρ 2 + P ( ρ ) ) − ( 1 − ρ 2 ) ( C D a ( ρ ) + 4 1 ) = ( f ( A ) − D a + 4 5 ) 1 − ρ 2 + P ( ρ ) ( ∵ C D a ( ρ ) = D a + 1 )
≥ 0.05 1 − ρ 2 + P ( ρ ) ( ∵ f ~ ( A ) ≥ f ~ ( 1.968 ) > 2.45 ) . \ge0.05\sqrt{1-\rho^2}+P(\rho)
\qquad\left(\because\widetilde{f}(A)\ge\widetilde{f}(1.968)>2.45\right). ≥ 0.05 1 − ρ 2 + P ( ρ ) ( ∵ f ( A ) ≥ f ( 1.968 ) > 2.45 ) .
Here, P ( ρ ) P(\rho) P ( ρ ) is a quadratic function that is increasing for all ρ ≥ 0.7 \rho\ge0.7 ρ ≥ 0.7 . Hence:
For ρ ∈ [ 0.7 , 0.735 ] \rho\in[0.7,0.735] ρ ∈ [ 0.7 , 0.735 ] , since P ( ρ ) ≥ P ( 0.7 ) > − 0.0288 P(\rho)\ge P(0.7)>-0.0288 P ( ρ ) ≥ P ( 0.7 ) > − 0.0288 and 1 − ρ 2 ≥ 1 − 0.735 2 > 0.678 \sqrt{1-\rho^2}\ge\sqrt{1-0.735^2}>0.678 1 − ρ 2 ≥ 1 − 0.73 5 2 > 0.678 , we have f A ( ρ ) − q c ( ρ ) > 0.05 ⋅ 0.678 − 0.0288 = 0.0051 > 0 f_A(\rho)-q_c(\rho)>0.05\cdot0.678-0.0288=0.0051>0 f A ( ρ ) − q c ( ρ ) > 0.05 ⋅ 0.678 − 0.0288 = 0.0051 > 0 .
For ρ ∈ [ 0.735 , 1 ] \rho\in[0.735,1] ρ ∈ [ 0.735 , 1 ] , since P ( ρ ) ≥ P ( 0.735 ) > 0 P(\rho)\ge P(0.735)>0 P ( ρ ) ≥ P ( 0.735 ) > 0 and 1 − ρ 2 ≥ 0 \sqrt{1-\rho^2}\ge0 1 − ρ 2 ≥ 0 , we have f A ( ρ ) − q c ( ρ ) > 0 f_A(\rho)-q_c(\rho)>0 f A ( ρ ) − q c ( ρ ) > 0 .
Therefore, Theorem 61 follows. □
Lemma 62. Suppose − 1 2 ≤ ρ ≤ 1 -\frac{1}{2}\le\rho\le1 − 2 1 ≤ ρ ≤ 1 . If 2.15 ≤ A ≤ A 0 2.15\le A\le A_0 2.15 ≤ A ≤ A 0 , then q A ( ρ ) = q c ( ρ ) q_A(\rho)=q_c(\rho) q A ( ρ ) = q c ( ρ ) , that is, f A ( ρ ) ≥ q c ( ρ ) f_A(\rho)\ge q_c(\rho) f A ( ρ ) ≥ q c ( ρ ) . This also proves property (P2).
Proof. Since f A ( ρ ) f_A(\rho) f A ( ρ ) is an increasing function with respect to A A A by property (P1), it suffices to prove for the case A = 2.15 A=2.15 A = 2.15 . The graph of f 2.15 ( ρ ) − q c ( ρ ) f_{2.15}(\rho)-q_c(\rho) f 2.15 ( ρ ) − q c ( ρ ) is presented in Figure 7 , which illustrates that this value is positive for all ρ \rho ρ . However, to prove that f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 f_{2.15}(\rho)-q_c(\rho)\ge0 f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 in a rigorous way, we use a known technique called interval arithmetic .
Figure 7. The graph of f 2.15 ( ρ ) − q c ( ρ ) f_{2.15}(\rho)-q_c(\rho) f 2.15 ( ρ ) − q c ( ρ ) for − 1 2 ≤ ρ ≤ 1 -\frac{1}{2}\leq\rho\leq1 − 2 1 ≤ ρ ≤ 1 . Note that the vertical axis is logarithmic.
Interval arithmetic. Interval arithmetic calculates a possible range of an expression given possible ranges of each value. For example, given that x ∈ [ − 0.5 , 1.2 ] x\in[-0.5,1.2] x ∈ [ − 0.5 , 1.2 ] , we have x 2 ∈ [ 0 2 , max ( 0.5 , 1.2 ) 2 ] = [ 0 , 1.44 ] x^2\in[0^2,\max(0.5,1.2)^2]=[0,1.44] x 2 ∈ [ 0 2 , max ( 0.5 , 1.2 ) 2 ] = [ 0 , 1.44 ] , and hence x 2 + x ∈ [ 0 − 0.5 , 1.44 + 1.2 ] = [ − 0.5 , 2.64 ] x^2+x\in[0-0.5,1.44+1.2]=[-0.5,2.64] x 2 + x ∈ [ 0 − 0.5 , 1.44 + 1.2 ] = [ − 0.5 , 2.64 ] . Note that the calculated range is not always tight: in fact, x 2 + x x^2+x x 2 + x is actually in the range [ − 0.25 , 2.64 ] [-0.25,2.64] [ − 0.25 , 2.64 ] . Interval arithmetic supports addition, subtraction, multiplication, division, and square root, so we can compute a plausible range for functions built from these elementary operations.
In our program, to avoid floating-point operations, we round both endpoints (i.e., round down the left and round up the right) to multiples of 2 − 72 2^{-72} 2 − 72 .
Proving the inequality with interval arithmetic. For a sufficiently small range [ ℓ , r ] [\ell,r] [ ℓ , r ] , we can prove that “f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 f_{2.15}(\rho)-q_c(\rho)\ge0 f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 for all ρ ∈ [ ℓ , r ] \rho\in[\ell,r] ρ ∈ [ ℓ , r ] ” by calculating a possible range of f 2.15 ( ρ ) − q c ( ρ ) f_{2.15}(\rho)-q_c(\rho) f 2.15 ( ρ ) − q c ( ρ ) with interval arithmetic, if the lower endpoint is positive.
To prove f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 f_{2.15}(\rho)-q_c(\rho)\ge0 f 2.15 ( ρ ) − q c ( ρ ) ≥ 0 for all ρ ∈ [ − 1 2 , 1 ] \rho\in[-\frac{1}{2},1] ρ ∈ [ − 2 1 , 1 ] , we divided the interval [ − 1 2 , 1 ] [-\frac{1}{2},1] [ − 2 1 , 1 ] into sufficiently small intervals [ ℓ , r ] [\ell,r] [ ℓ , r ] , so that the inequality is proved inside these smaller intervals. We note that the case ρ ≥ 7 10 \rho\ge\frac{7}{10} ρ ≥ 10 7 is already proved in Theorem 61. We provide a complete certificate and a verifier in the supplementary material. □
Therefore, it remains to prove Case 1 for the range ( A , ρ ) ∈ [ 1.968 , 2.15 ] × [ − 1 2 , 0.7 ] (A,\rho)\in[1.968,2.15]\times[-\frac{1}{2},0.7] ( A , ρ ) ∈ [ 1.968 , 2.15 ] × [ − 2 1 , 0.7 ] and Case 2 for all A ∈ [ 1.968 , A 0 ] A\in[1.968,A_0] A ∈ [ 1.968 , A 0 ] . This is proved as follows: we provide a complete certificate and a verifier in the supplementary material.
Proof of Theorem 40. Here, we explain how to prove Case 1, since Case 2 can be proved similarly (and more easily). As in the proof of Theorem 62, we divide the rectangle [ 1.968 , 2.15 ] × [ − 1 2 , 0.7 ] [1.968,2.15]\times[-\frac{1}{2},0.7] [ 1.968 , 2.15 ] × [ − 2 1 , 0.7 ] into sufficiently small subrectangles [ A ℓ , A r ] × [ ρ ℓ , ρ r ] [A_\ell,A_r]\times[\rho_\ell,\rho_r] [ A ℓ , A r ] × [ ρ ℓ , ρ r ] , and for each of them, we prove Theorem 40 for all ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] (A,\rho)\in[A_\ell,A_r]\times[\rho_\ell,\rho_r] ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] . Specifically, for each [ A ℓ , A r ] × [ ρ ℓ , ρ r ] [A_\ell,A_r]\times[\rho_\ell,\rho_r] [ A ℓ , A r ] × [ ρ ℓ , ρ r ] we show either of the following:
Prove that q A ( ρ ) = q c ( ρ ) q_A(\rho)=q_c(\rho) q A ( ρ ) = q c ( ρ ) for all ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] (A,\rho)\in[A_\ell,A_r]\times[\rho_\ell,\rho_r] ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] , which means that the assumption of Case 1 does not hold. This can be done very similarly to the proof of Theorem 62.
Give one tuple ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) such that the conditions (i), (ii), (iii), and (iv) hold for all ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] (A,\rho)\in[A_\ell,A_r]\times[\rho_\ell,\rho_r] ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] .
Next, we explain how to prove conditions (i), (ii), (iii), and (iv) for a given ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) . The verification of conditions (i) and (ii) is trivial. For condition (iii), we need to show that
( q A ( ρ ) 1 − ρ 2 ) 2 + ( μ + θ ( μ + κ ) ) 2 − ( C D a ( ρ ) + 1 4 ) 2 − 10 − 18 ≥ 0 (8) \left(\frac{q_A(\rho)}{\sqrt{1-\rho^2}}\right)^2+(\mu+\theta(\mu+\kappa))^2-\left(C_{\mathrm{D_a}}(\rho)+\frac{1}{4}\right)^2-10^{-18}\geq0 \tag*{(8)} ( 1 − ρ 2 q A ( ρ ) ) 2 + ( μ + θ ( μ + κ ) ) 2 − ( C D a ( ρ ) + 4 1 ) 2 − 1 0 − 18 ≥ 0 (8)
for all ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] (A,\rho)\in[A_\ell,A_r]\times[\rho_\ell,\rho_r] ( A , ρ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] . The verification of condition (iv) is trickier: we show that
L ( 1 / 2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ / 2 ) + θ c ∣ ρ ′ + ρ / 2 ∣ 1 − ρ 2 + 3 2 ( 1 + θ ) a − 3 2 θ A − q A ( ρ ′ ) − 3 ⋅ 10 − 6 ≥ 0 (9) \frac{L(1/2+\rho\rho')-(1+\theta)b(\rho'+\rho/2)+\theta c\lvert\rho'+\rho/2\rvert}{\sqrt{1-\rho^2}}+\frac{3}{2}(1+\theta)a-\frac{\sqrt{3}}{2}\theta A-q_A(\rho')-3\cdot10^{-6}\geq0 \tag*{(9)} 1 − ρ 2 L ( 1/2 + ρ ρ ′ ) − ( 1 + θ ) b ( ρ ′ + ρ /2 ) + θ c ∣ ρ ′ + ρ /2 ∣ + 2 3 ( 1 + θ ) a − 2 3 θ A − q A ( ρ ′ ) − 3 ⋅ 1 0 − 6 ≥ 0 (9)
where L = q A ( ρ ) / 1 − ρ 2 L=q_A(\rho)/\sqrt{1-\rho^2} L = q A ( ρ ) / 1 − ρ 2 , for all ( A , ρ , ρ ′ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] × [ − 1 2 , 1 2 − ρ ℓ ] (A,\rho,\rho')\in[A_\ell,A_r]\times[\rho_\ell,\rho_r]\times[-\frac{1}{2},\frac{1}{2}-\rho_\ell] ( A , ρ , ρ ′ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] × [ − 2 1 , 2 1 − ρ ℓ ] . Here, we need to prove this over the three-dimensional box (rather than the two-dimensional rectangle).
For condition (iv), we further divide ρ ′ ∈ [ − 1 2 , 1 2 − ρ ℓ ] \rho' \in[-\frac{1}{2}, \frac{1}{2}-\rho_\ell] ρ ′ ∈ [ − 2 1 , 2 1 − ρ ℓ ] into multiple intervals [ ρ ℓ ′ , ρ r ′ ] [\rho'_\ell,\rho'_r] [ ρ ℓ ′ , ρ r ′ ] , and for each of them, we prove the inequality for all ( A , ρ , ρ ′ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] × [ ρ ℓ ′ , ρ r ′ ] (A,\rho,\rho') \in[A_\ell,A_r] \times[\rho_\ell,\rho_r] \times[\rho'_\ell,\rho'_r] ( A , ρ , ρ ′ ) ∈ [ A ℓ , A r ] × [ ρ ℓ , ρ r ] × [ ρ ℓ ′ , ρ r ′ ] . In order to guarantee differentiability, for each [ ρ ℓ ′ , ρ r ′ ] [\rho'_\ell,\rho'_r] [ ρ ℓ ′ , ρ r ′ ] , we prove an inequality that replaces ∣ ρ ′ + ρ / 2 ∣ |\rho' + \rho/2| ∣ ρ ′ + ρ /2∣ with either + ( ρ ′ + ρ / 2 ) +(\rho' + \rho/2) + ( ρ ′ + ρ /2 ) or − ( ρ ′ + ρ / 2 ) -(\rho' + \rho/2) − ( ρ ′ + ρ /2 ) (which sign to choose can differ for each [ ρ ℓ ′ , ρ r ′ ] [\rho'_\ell,\rho'_r] [ ρ ℓ ′ , ρ r ′ ] ); this is indeed a stronger inequality, due to θ > 0 \theta> 0 θ > 0 , c ≥ 0 c \ge0 c ≥ 0 , and 1 − ρ 2 ≥ 0 \sqrt{1-\rho^2} \ge0 1 − ρ 2 ≥ 0 . Similarly, we also replace q A ( ρ ′ ) q_A(\rho') q A ( ρ ′ ) with either f A ( ρ ′ ) f_A(\rho') f A ( ρ ′ ) or q C ( ρ ′ ) q_C(\rho') q C ( ρ ′ ) ; this only makes the inequality stronger.
Differentiability of the functions. Generally speaking, let F : R d → R F : \mathbb{R}^d \to\mathbb{R} F : R d → R be a multivariable function composed of addition, subtraction, multiplication, division, and square root. The function is C ∞ C^\infty C ∞ (i.e., infinitely differentiable) on any open region where every denominator of division is non-zero and every argument of square root is positive.
Therefore, the left-hand sides of the two inequalities above (for conditions (iii) and (iv); Equation (8) and Equation (9) ) are C ∞ C^\infty C ∞ for ρ , ρ ′ ≠ − 1 40 , r − , 0 , r + , 1 50 , 1 4 , 1 \rho,\rho' \ne-\frac{1}{40}, r_-, 0, r_+, \frac{1}{50}, \frac{1}{4}, 1 ρ , ρ ′ = − 40 1 , r − , 0 , r + , 50 1 , 4 1 , 1 , where
r − : = 4 − D a 2 D a + 16 ≈ − 0.020238 , r + : = D a − 4 4 D a + 8 ≈ 0.019451. r_- := \frac{4-D_a}{2D_a+16} \approx-0.020238,\qquad r_+ := \frac{D_a-4}{4D_a+8} \approx0.019451. r − := 2 D a + 16 4 − D a ≈ − 0.020238 , r + := 4 D a + 8 D a − 4 ≈ 0.019451.
These are the points where the capacity C D ( ρ ) C_D(\rho) C D ( ρ ) in Theorem 36 matches the threshold D + 2 D+2 D + 2 . Moreover, for condition (iii), if we divide the rectangle [ 1.968 , 2.15 ] × [ − 1 2 , 0.7 ] [1.968,2.15] \times[-\frac{1}{2},0.7] [ 1.968 , 2.15 ] × [ − 2 1 , 0.7 ] into seven subrectangles splitting at ρ = − 1 40 , r − , 0 , r + , 1 50 , 1 4 \rho= -\frac{1}{40},r_-,0,r_+,\frac{1}{50},\frac{1}{4} ρ = − 40 1 , r − , 0 , r + , 50 1 , 4 1 , the left-hand side of Equation (8) can be considered C ∞ C^\infty C ∞ for each box, including the boundary, because this eliminates case divisions (including min and max). Similarly, for condition (iv), if we divide the box so that its interior does not contain ρ ′ = − 1 40 , r − , 0 , r + , 1 50 , 1 4 \rho' = -\frac{1}{40},r_-,0,r_+,\frac{1}{50},\frac{1}{4} ρ ′ = − 40 1 , r − , 0 , r + , 50 1 , 4 1 , the left-hand side of Equation (9) can be considered C ∞ C^\infty C ∞ for each box (except for ρ ′ = 1 \rho'=1 ρ ′ = 1 ).
Proving the positivity of a function. In general, we need to prove that a C ∞ C^\infty C ∞ function F : D → R F : D \to\mathbb{R} F : D → R is positive inside the domain D D D , where D D D is a rectangle for condition (iii) and a box for condition (iv). For each region, we choose either of the following ways to prove the positivity.
Use interval arithmetic directly to compute the lower (and upper) bounds of F F F in the region, as explained in the proof of Theorem 62.
Apply a Taylor expansion to the function F F F and obtain a lower bound explained below.
If the resulting lower bound is positive, it proves that F F F is positive.
Taylor expansion. Here, we consider the two-dimensional case where D = [ x ℓ , x r ] × [ y ℓ , y r ] D = [x_\ell,x_r] \times[y_\ell,y_r] D = [ x ℓ , x r ] × [ y ℓ , y r ] ; the three-dimensional case can be handled similarly. We define
x m : = x ℓ + x r 2 , y m : = y ℓ + y r 2 x_m := \frac{x_\ell+x_r}{2},\qquad y_m := \frac{y_\ell+y_r}{2} x m := 2 x ℓ + x r , y m := 2 y ℓ + y r
for the center. Then, for ( x , y ) ∈ D (x,y) \in D ( x , y ) ∈ D , by Taylor’s theorem, there exists some ( x ′ , y ′ ) ∈ D (x',y') \in D ( x ′ , y ′ ) ∈ D such that:
f ( x , y ) = f ( x m , y m ) + ∂ f ∂ x ( x m , y m ) ⋅ ( x − x m ) + ∂ f ∂ y ( x m , y m ) ⋅ ( y − y m ) + 1 2 ∂ 2 f ∂ x 2 ( x ′ , y ′ ) ⋅ ( x − x m ) 2 + 1 2 ∂ 2 f ∂ y 2 ( x ′ , y ′ ) ⋅ ( y − y m ) 2 + ∂ 2 f ∂ x ∂ y ( x ′ , y ′ ) ⋅ ( x − x m ) ( y − y m ) . \begin{aligned}
f(x,y) ={}& f(x_m,y_m) + \frac{\partial f}{\partial x}(x_m,y_m)\cdot(x-x_m) + \frac{\partial f}{\partial y}(x_m,y_m)\cdot(y-y_m) \\
&+ \frac{1}{2}\frac{\partial^2 f}{\partial x^2}(x',y')\cdot(x-x_m)^2 + \frac{1}{2}\frac{\partial^2 f}{\partial y^2}(x',y')\cdot(y-y_m)^2 \\
&+ \frac{\partial^2 f}{\partial x\partial y}(x',y')\cdot(x-x_m)(y-y_m).
\end{aligned} f ( x , y ) = f ( x m , y m ) + ∂ x ∂ f ( x m , y m ) ⋅ ( x − x m ) + ∂ y ∂ f ( x m , y m ) ⋅ ( y − y m ) + 2 1 ∂ x 2 ∂ 2 f ( x ′ , y ′ ) ⋅ ( x − x m ) 2 + 2 1 ∂ y 2 ∂ 2 f ( x ′ , y ′ ) ⋅ ( y − y m ) 2 + ∂ x ∂ y ∂ 2 f ( x ′ , y ′ ) ⋅ ( x − x m ) ( y − y m ) .
Therefore, a lower bound on the minimum value of f ( x , y ) f(x,y) f ( x , y ) can be obtained using
min ( x ′ , y ′ ) ∈ D ∂ 2 f ∂ x 2 ( x ′ , y ′ ) , min ( x ′ , y ′ ) ∈ D ∂ 2 f ∂ y 2 ( x ′ , y ′ ) , max ( x ′ , y ′ ) ∈ D ∣ ∂ 2 f ∂ x ∂ y ( x ′ , y ′ ) ∣ , \min_{(x',y')\in D} \frac{\partial^2 f}{\partial x^2}(x',y'), \qquad\min_{(x',y')\in D} \frac{\partial^2 f}{\partial y^2}(x',y'), \qquad\max_{(x',y')\in D} \left|\frac{\partial^2 f}{\partial x\partial y}(x',y')\right|, ( x ′ , y ′ ) ∈ D min ∂ x 2 ∂ 2 f ( x ′ , y ′ ) , ( x ′ , y ′ ) ∈ D min ∂ y 2 ∂ 2 f ( x ′ , y ′ ) , ( x ′ , y ′ ) ∈ D max ∂ x ∂ y ∂ 2 f ( x ′ , y ′ ) ,
where the bounds on these three values are calculated by interval arithmetic.
Concluding the proof. With the methods explained above, for all ( A , ρ ) (A,\rho) ( A , ρ ) that meet the assumption of Case 1, we proved the existence of a tuple ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) that satisfies all of conditions (i), (ii), (iii), and (iv). We also proved this similarly for Case 2. A complete certificate and a verifier is provided at https://github.com/ka714/coloring_01537 . □ \square □
Remark 63. Although we explained the proof of the Theorem 40 in the discussion above, we did not explain how to obtain a certificate, that is, how to divide a rectangle [ 1.968 , 2.15 ] × [ − 1 2 , 0.7 ] [1.968, 2.15] \times[-\frac{1}{2}, 0.7] [ 1.968 , 2.15 ] × [ − 2 1 , 0.7 ] , so we explain it here. We start with the entire [ 1.968 , 2.15 ] × [ ρ ℓ , ρ r ] [1.968, 2.15] \times[\rho_\ell, \rho_r] [ 1.968 , 2.15 ] × [ ρ ℓ , ρ r ] (which is one of the seven rectangles mentioned in the paragraph “Differentiability of the Functions”) and attempt to find a common ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) that meets conditions (i), (ii), (iii), and (iv) in the entire domain. If no such tuple ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) can be found, we divide the rectangle into two equal parts, either horizontally or vertically, and recursively apply the same procedure. If we recurse deep enough and the rectangle becomes sufficiently small, we can find a tuple ( a , b , μ , c , κ , θ ) (a,b,\mu,c,\kappa,\theta) ( a , b , μ , c , κ , θ ) that satisfies the conditions. In this way, we obtain a finite-size certificate that covers the entire rectangle.