Some Lower Bounds on the Reach of an Algebraic Variety

Chris La Valle Josué Tonelli-Cueto

The University of Texas at San Antonio The Johns Hopkins University
Department of Mathematics Department of Applied Mathematics and Statistics
San Antonio, Texas, USA Baltimore, Maryland, USA
christopher.lavalle@my.utsa.edu josue.tonelli.cueto@bizkaia.eu
arXiv:2402.15649v1 [math.AG] 23 Feb 2024

ABSTRACT 1 INTRODUCTION
Separation bounds are a fundamental measure of the complexity of The reach [18] is a positive dimensional generalization of the sep-
solving a zero-dimensional system as it measures how difficult it is aration between zeros. This quantity is fundamental in the homol-
to separate its zeroes. In the positive dimensional case, the notion ogy inference for submanifolds [22, 23], and, in particular, for al-
of reach takes its place. In this paper, we provide bounds on the gebraic and semialgebraic sets [4–7]. Recently, a large amount of
reach of a smooth algebraic variety in terms of several invariants interest has emerged in the study of the reach of algebraic vari-
of interest: the condition number, Smale’s 𝛾 and the bit-size. We eties [2, 9–12, 19, 20], but, aside from a few exceptions [4, 7], the
also provide probabilistic bounds for random algebraic varieties existing lower bounds for the reach are scarce.
under some general assumptions. In this paper, we aim to close this gap in the literature. Building
off the work of [4, 7], [16, 17, 28], [15, 26, 27] and [21], we improve
KEYWORDS some of the existing lower bounds for the reach of an algebraic
condition number; real algebraic geometry; random algebraic ge- variety, providing new probabilistic bounds. As the reach can be
ometry; reach seen as a positive dimensional analogue of the separation of the
zeros, this can be seen as a generalization and expansion of [13, 14]
ACM Reference Format:
Chris La Valle and Josué Tonelli-Cueto. 2024. Some Lower Bounds on the to the positive dimensional case.
Reach of an Algebraic Variety. In Proceedings of . ACM, New York, NY, USA,
9 pages. https://doi.org/10.1145/nnnnnnn.nnnnnnn Notations
P𝑛,𝑑 [𝑞] will denote the set of real polynomial 𝑞-tuples, where the
𝑖th polynomial has degree ≤ 𝑑𝑖 . For a polynomial tuple 𝑓 ∈ P𝑛,𝑑 [𝑞]
and 𝑥 ∈ R𝑛 , Z(𝑓 ) denotes the real zero, D𝑥 𝑓 the tangent map of
𝑓 at 𝑥, and D𝑘𝑥 𝑓 the 𝑘th derivative tensor of 𝑓 . We will use 𝜕𝑘 𝑓
to denote the corresponding tensor of polynomials formed by the
partial derivatives of 𝑓 , and 𝜕𝑘 𝑓 [𝑣 1 , . . . , 𝑣𝑘 ] to denote the polyno-
mial obtained when contracting the polynomial tensor 𝜕𝑘 𝑓 with
Î
the vectors 𝑣 1 , . . . , 𝑣𝑘 . We will denote random variables as 𝔣, 𝔳 . . .
k𝐴k𝑟,𝑠 := sup𝑣1,...,𝑣ℓ k𝐴(𝑣 1, . . . , 𝑣 ℓ )k𝑠 / 𝑖 k𝑣𝑖 k𝑟 will denote the in-
duced norm for a multilinear map 𝐴 : (R𝑎 ) ℓ → R𝑏 .

2 MAIN RESULTS
The medial axis of a closed set 𝑍 ⊆ R𝑛 , M (𝑋 ), is the set of those
points with more than one nearest point ot 𝑋 , i.e.,

M (𝑍 ) := 𝑝 ∈ R𝑛 | #{𝜁 ∈ 𝑍 | dist2 (𝑝, 𝜁 ) = dist2 (𝑝, 𝑍 )} > 1 .
(2.1)
The distance from 𝑍 to its medial axis is the reach of 𝑍 , 𝜚 (𝑍 ), i.e.,
𝜚 (𝑍 ) := dist2 (𝑍, M (𝑍 )) = min{dist2 (𝜁 , 𝑝) | (𝜁 , 𝑝) ∈ 𝑍 × M (𝑍 )};
(2.2)
and the distance from 𝜁 ∈ 𝑍 to the medial axis of 𝑍 is the local
reach of 𝑍 at 𝜁 , 𝜚 (𝑍, 𝜁 ), i.e.,
𝜚 (𝑍, 𝜁 ) := dist2 (𝜁 , M (𝑍 )) = min{dist2 (𝜁 , 𝑝) | 𝑝 ∈ M (𝑍 )}. (2.3)
The reach is in this way a higher dimensional analogue of separa-
tion, we provide lower bounds for the following related quantity:
𝜚 𝑅 (𝑍 ) := min{𝜚 (𝑍, 𝜁 ) | 𝜁 ∈ 𝑍, k𝜁 k ∞ ≤ 𝑅}. (2.4)
We provide bounds in three cases:
,, Chris La Valle and Josué Tonelli-Cueto

Theorem 2.1. Let 𝑅 ∈ N and 𝑓 ∈ P𝑛,𝑑 [𝑞] be a tuple of integer avoid confusion, we use 𝜕𝑓 [𝑣] to denote the 𝑞-tuple of polynomi-
polynomials with coefficients of bit-size at most 𝜏. Then 𝜚 𝑅 (Z(𝑓 )) = als given by D𝑋 𝑓 [𝑣], reserving the symbol D to denote the Jaco-
0 or log 1/𝜚 𝑅 (Z(𝑓 )) is bounded by bian of 𝑓 (and higher derivatives) at a particular point. In other
words, we use the symbol 𝜕 to emphasize that we are considering
4𝑛(2D) 1+𝑞+2𝑛 (5 + 𝜏 + log 𝑅 + 6𝑛 log D) + 2 log D + 𝜏.
the polynomial.
Theorem 2.2. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a random integer polynomial
Theorem 4.1. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and 𝑣 ∈ R𝑛 . Then
h
whose coefficients are independent and uniformly distributed in Z ∩ i


[−2𝜏 , 2𝜏 ]. Then for 𝑡 ∈ max{log D, log 𝑛}, log 𝑞(𝑛 + 1) 𝑛+D Δ −1 𝜕𝑓 𝑣 1 ≤ k 𝑓 k 1 k𝑣 k ∞ (4.2)
𝑛 +𝜏 −1 ,
     where Δ = diag(𝒅)
1 𝑛
+1
√ 𝑛 + D 𝑞+𝑛 𝑞+2𝑛 −𝑡
P log ≥ 𝑡 ≤ 20𝑅𝑛 2 2 D 2 . The following corollaries follow from applying iteratively the
𝜚 𝑅 (Z(𝔣)) 𝑛
theorem above. Recall that for a ℓ-multilinear map 𝐴 : R𝑛1 ×· · · R𝑛ℓ →
Theorem 2.3. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a random polynomial tuple R𝑚 , we define its (𝑝, 𝑞)-spectral norm, k k 𝑝,𝑞 , by
whose coefficients are independent and uniformly distributed in [−1, 1].
Then for 𝑡 ≥ 2(𝑛 + 1), k𝐴[𝑣 1, . . . , 𝑣 ℓ ] k𝑞
k𝐴k 𝑝,𝑞 := sup (4.3)
   𝑣1,...,𝑣ℓ ≠0 k𝑣 1 k 𝑝 · · · k𝑣 ℓ k 𝑝
1 𝑛 𝑛 + D 𝑛+𝑞 𝑛+𝑞
P(log ≥ 𝑡) ≤ 8𝑅(𝑛+1) 2 +1 D𝑞+2𝑛 4 𝑡 𝑝 2 −𝑡 . where 𝐴[𝑣 1 , . . . , 𝑣 ℓ ] is the evaluation of 𝐴 at (𝑣 1, . . . , 𝑣𝑘 ). Following
𝜚 𝑅 (Z(𝔣)) 𝑛
the convention before, we denote by 𝜕 ℓ 𝑓 [𝑣 1 , . . . , 𝑣 ℓ ] the 𝑞-tuple of
The first theorem provides a worst case bound on the reach of polynomials D𝑋 ℓ 𝑓 [𝑣 , . . . , 𝑣 ].
1 ℓ
an algebraic variety given by integer polynomials. The second and
third provide probabilistic bounds for the reach in the case of a Corollary 4.2. Let 𝑓 ∈ P𝑛,𝑑 [𝑞], ℓ ∈ N and 𝑣 1 , . . . , 𝑣 ℓ ∈ R𝑛 .
 
random polynomial with, respectively, random integer coefficients Then
and random continuous real coefficients. The probabilistic theo- 1 ℓ D
𝜕 𝑓 [𝑣 1 , . . . , 𝑣 ℓ ] ≤ k 𝑓 k 1 k𝑣 1 k ∞ · · · k𝑣 ℓ k ∞ . (4.4)
rems are, respectively, particular cases of Theorems 8.8 and 7.6. ℓ! 1 ℓ
Moreover, we provide bounds of the reach in terms of the con- Moreover,
dition number (Theorem 6.6), Kantorovich’s theory (Theorem 5.2)  
1 𝑘 Δ
and Smale’s 𝛼-theory (Theorem 5.5). 𝜕 𝑓 [𝑣 1 , . . . , 𝑣 ℓ ] ≤ 𝑓 k𝑣 1 k ∞ · · · k𝑣 ℓ k ∞ (4.5)
ℓ! ℓ


1 1
3 FEDERER’S ESTIMATORS OF THE REACH where Δℓ := Δ(Δ − I) · · · (Δ − (ℓ − 1)I)/ℓ!.
The following results can be found in some form in [18]. As of to-
Corollary 4.3. Let 𝑓 ∈ P𝑛,𝑑 [𝑞], 𝑥 ∈ R𝑛 and ℓ ∈ N. Then
day, they are fundamental in the estimation of the reach of mani-  
folds, since they allow us to compute the reach. Moreover, the most 1 ℓ D
𝜕𝑥 𝑓 ≤ k 𝑓 k 1 max{1, k𝑥 k D−ℓ }. (4.6)
widely used estimator for the reach [1] uses them. ℓ! ∞,∞ ℓ
Theorem 3.1. Let 𝑍 ⊆ R𝑚 be a closed set, 𝜁 ∈ 𝑍 and 𝑟, 𝑡 > 0. If Moreover,  
𝑍 ∩ 𝐵 (𝜁 , 𝑟 ) is a closed submanifold of 𝐵 (𝜁 , 𝑟 ) and for all 𝑧, ˜𝑧∈𝑍∩ −𝒅+ℓ1 1 ℓ Δ
k𝑥 k h∞ D 𝑓 ≤ 𝑓 (4.7)
𝐵 (𝜁 , 𝑟 ), dist2 (𝑧˜ − 𝑧, T𝑧 𝑍 ) ≤ k𝑧˜ − 𝑧 k 22 /(2𝑡), then 𝜚 (𝑍, 𝜁 ) ≥ min{𝑟, 𝑡 }. ℓ! 𝑥 ℓ


∞,∞ 1

Corollary 3.2. Let 𝑍 ⊆ R𝑚 be a locally closed submanifold. where Δℓ as in Corollary 4.2 and k𝑥 k h∞
𝒅−ℓ1 := diag(max{1, k𝑥 k}𝑑𝑖 −ℓ ).

Then Corollary 4.4. Let 𝑓 ∈ P𝑛,𝑑 [𝑞], 𝑘 ∈ N and 𝑣 1, . . . , 𝑣𝑘 ∈ R𝑛 .

k𝑧˜ − 𝑧 k 22 Then the function
𝜏 (𝑍 ) = inf .
𝑧,𝑧˜ ∈𝑍 2dist2 (𝑧˜ − 𝑧, T𝑧 𝑍 ) 1
𝐼𝑛 ∋ 𝑥 ↦→ D𝑥ℓ 𝑓 [𝑣 1 /k𝑣 1 k ∞, . . . , 𝑣 ℓ /k𝑣 ℓ k ∞ ] ∈ R𝑞
Proof of Theorem 3.1. This is a variation of [18, Lemma 4.17]. ℓ!
To see this variation, we only have to take 𝑥, 𝑎, 𝑏 ∈ 𝐵 (𝜁 , 𝑟 ) as in the Δ−I

is Δ ℓ 𝑓 -Lipschitz with respect the ∞-norm.
proof of [18, Lemma 4.17], i.e., 𝑎, 𝑏 ∈ 𝑍 and 𝑥 such that dist2 (𝑥, 𝑍 ) = 1
k𝑎 − 𝑥 k 2 = k𝑏 − 𝑥 k 2 < min{𝑟, 𝑡 }. Then the argument in the proof Proof of Theorem 4.1. Without loss of generality k𝑣 k ∞ = 1.
gives that 𝑎 = 𝑏. Hence dist2 (Δ(𝑍 ), 𝑍 ∩ 𝐵 (𝜁 , 𝑟 )) ≥ min{𝑟, 𝑡 } and so Í𝑛 𝜕𝑓𝑖 Í 𝜕𝑓
For any 𝑖, k𝜕𝑓𝑖 (𝑣)k 1 = 𝑣
𝑘=1 𝜕𝑋𝑘 𝑘 1
≤ 𝑛𝑘=1 𝜕𝑋𝑖𝑘 . Now,
Í
𝜚 (𝑋 , 𝜁 ) ≥ min{𝑟, 𝑡 }.  1
𝜕𝑓𝑖
𝜕𝑋𝑘 1 = 𝛼 𝛼𝑘 |𝑓𝑖,𝛼 |, thus
4 1-NORM ON POLYNOMIALS !
Í
𝑞 Õ 𝑛
𝜕𝑓𝑖 Õ𝑛 Õ Õ Õ 𝑛
Given a 𝑞-tuple of polynomials 𝑓 = 𝛼 𝑓𝑖,𝛼 𝑋 𝛼 𝑖=1 ∈ P𝑛,𝑑 [𝑞], we = 𝛼𝑘 |𝑓𝑖,𝛼 | ≤ 𝛼𝑘 |𝑓𝑖,𝛼 | ≤ 𝑑𝑖 k 𝑓𝑖 k 1
consider the 1-norm k k 1 given by 𝜕𝑋𝑘 1
Õ
𝑘=1 𝛼 𝛼 𝑘=1
Í
𝑘=1

k 𝑓 k 1 := max |𝑓𝑖,𝛼 |. (4.1) since the degree of 𝑓𝑖 being at most 𝑑𝑖 implies 𝑛𝑘=1 𝛼𝑘 ≤ 𝑑𝑖 . Hence
𝑖
𝛼
Δ −1 𝜕𝑓 𝑣 1 = max 𝑑𝑖−1 𝜕𝑓𝑖 𝑣 1 ≤ max k 𝑓𝑖 k 1 = k 𝑓 k 1,
The most important theorem regarding these two norms is the fol- 𝑖 𝑖
lowing theorem relating the norm of 𝑓 and that of 𝛿D𝑋 𝑓 [𝑣]. To as desired. 
Some Lower Bounds on the Reach of an Algebraic Variety ,,

Proof of Corollary 4.2. We only need to prove the second fam- and for some 𝑦 ∈ [𝑧, 𝜁 ],
ily of inequalities. Note that, by substituting 𝑓 by Δ𝑓 , we can rewrite
D𝜁 𝑓 † (D𝑧 𝑓 − D𝜁 𝑓 ) = D𝜁 𝑓 † D2𝑦 𝑓 [𝑧 − 𝜁 ]. (5.4)
Theorem 4.1 as k𝛿 𝑓 [𝑣] k 1 ≤ kΔ𝑓 k 1 . Now,
  Hence, [𝑧,
˜ 𝑧], [𝑧, 𝜁 ] ⊆ 𝐵 (𝜁 , 𝑟 ),
𝜕 ℓ 𝑓 [𝑣 1 , . . . , 𝑣 ℓ ] = 𝜕 𝜕 ℓ −1 𝑓 [𝑣 1 , . . . , 𝑣 ℓ −1 ] [𝑣 ℓ ]. (4.8) 1


kD𝜁 𝑓 † D𝑧 𝑓 [𝑧˜ − 𝑧] k 2 ≤
2
𝐾 (𝑓 , 𝜁 , 𝑟 )k𝑧˜ − 𝑧 k 22 (5.5)
Hence the claim follows by induction, since Δℓ = Δℓ Δ−I ℓ −1 . 
and
Proof of Corollary 4.3. This follows immediately from Corol- kD𝜁 𝑓 † (D𝑧 𝑓 − D𝜁 𝑓 )k 2,2 ≤ 𝐾 (𝑓 , 𝜁 , 𝑟 )𝑟 . (5.6)
lary 4.2 and an elementary bound.  The claim now follows combining these bounds, (5.2) and Theo-
rem 3.1, the claim follows. 
Proof of Corollary 4.4. By (4.8) and the mean value theorem,
we only need to show that Proof of Lemma 5.3. Since 𝐵 is surjective, then 𝐵 † is injective
   
1 ℓ+1 𝑣1 𝑣ℓ Δ−I and so the rank of 𝐴 and 𝐵 †𝐴 is the same. Now, let 𝜎𝑞 be the 𝑞th
𝜕 𝑓 .. ., ,𝑤 ≤ Δ 𝑓 k𝑤 k ∞ . singular value,
ℓ! k𝑣 1 k ∞, k𝑣 ℓ k ∞ ℓ 1
Now, this follows from Corollary 4.3.  𝜎𝑞 (𝐵 †𝐴) ≥ 𝜎𝑞 (𝐵 †𝐵) − k𝐵 † (𝐴 − 𝐵)k 2,2 = 1 − k𝐵 † (𝐴 − 𝐵)k 2,2 > 0
by the 1-Lipschitz property of 𝜎𝑞 , the fact that 𝐵 †𝐵 is an orthogo-
5 KANTOROVICH, SMALE AND REACH nal projection onto a 𝑞-dimensional subspace and the assumption.
Kantorovich’s theory and Smale’s 𝛼-theory are at the core of certi- Therefore rank 𝐵 †𝐴 ≥ 𝑞 and so rank 𝐴 ≥ 𝑞 and 𝐴 is surjective
fying fast convergence of Newton’s method. Now, they also appear therefore.
in bounds of the reach. For the inequality, note that, since 𝐵𝐵 † = I,
𝐴† 𝐵 = (𝐵 − (𝐴 − 𝐵)) † 𝐵 = (𝐵 (I − 𝐵 † (𝐴 − 𝐵))) † 𝐵.
Í∞
5.1 Kantorovich’s Theory and Reach
Kantorovich’s theory allows us to decide the fast convergence of Here, I−𝐵 † (𝐴−𝐵) is invertible with inverse I+ 𝑘=1 (−1)𝑘 −𝐵 † (𝐴−
Newton’s method for a 𝐶 2 -function using 𝑓 (𝑥), D𝑥 𝑓 and the values 𝐵) since k𝐵 † (𝐴 − 𝐵)k 2,2 < 1. Thus
of D𝑧2 𝑓 around 𝑥 (see [8, Théorème 88] among others). 𝐴† 𝐵 = (𝐵 (I − 𝐵 † (𝐴 − 𝐵))) † 𝐵 (I − 𝐵 † (𝐴 − 𝐵))(I − 𝐵 † (𝐴 − 𝐵)) −1 .
Definition 5.1. Let 𝑓 : R𝑛
→ beR𝑞 a 𝐶 2 -function
and 𝑥 ∈ R𝑛 , Hence, since (𝐵 (I − 𝐵 † (𝐴 − 𝐵))) † 𝐵 (I − 𝐵 † (𝐴 − 𝐵)) is an orthogonal
then 𝛽 (𝑓 , 𝑥) := kD𝑥 𝑓 † 𝑓 (𝑥)k 2 and Kantorovich’s regularity measure projection,
n o
is
k𝐴† 𝐵 k 2,2 ≤ k (I − 𝐵 † (𝐴 − 𝐵)) −1 k 2,2 ≤ 1/(1 − k𝐵 † (𝐴 − 𝐵)k 2,2 ),
𝐾 (𝑓 , 𝑥, 𝑟 ) := max kD𝑥 𝑓 † D𝑧2 𝑓 k 2,2 | k𝑧 − 𝑥 k 2 ≤ 𝑟 . (5.1)
by the series for the inverse. 
We give a lower bound of the local reach in terms of the Kan-
torovich’s regularity measure. 5.2 Smale’s 𝛼-theory and Reach
Smale’s 𝛼-theory allows us to decide the fast convergence of New-
Theorem 5.2. Let 𝑓 : R𝑛 → R𝑞 be a 𝐶 2 -function, 𝜁 ∈ Z(𝑓 ) :=
ton’s method for an analytic function with information depending
{𝑥 ∈ R𝑛 | 𝑓 (𝑥) = 0} and 𝑟 > 0. If D𝜁 𝑓 is surjective and 𝐾 (𝑓 , 𝜁 , 𝑟 )𝑟 <
on the value of 𝑓 and its derivatives at the initial point 𝑥 (see [8,
1, then
Théorème 128] among others).
𝜚 (Z(𝑓 ), 𝜁 ) ≥ 𝑟 .
In particular, if 1/2 ≤ 𝐾 (𝑓 , 𝜁 , 𝑟 )𝑟 ≤ 1, then 𝜚 (Z(𝑓 ), 𝜁 ) ≥ 1/(2𝐾 (𝑓 , 𝜁 , 𝑟 )). Definition 5.4. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and 𝑥 ∈ R𝑛 . Then:
(𝛼) Smale’s 𝛼: 𝛼 (𝑓 , 𝑥) := 𝛽 (𝑓 , 𝑥)𝛾 (𝑓 , 𝑥).
Lemma 5.3. Let 𝐴, 𝐵 ∈ R𝑞×𝑛 . If 𝐵 is surjective and k𝐵 † (𝐴−𝐵)k 2 < (𝛽) Smale’s 𝛽: 𝛽 (𝑓 , 𝑥) := kD𝑥 𝑓 † 𝑓 (𝑥)k 2 .
1, then 𝐴 is surjective and (𝛾) Smale’s 𝛾: 𝛾 (𝑓 , 𝑥) := supℓ ≥2 D𝑥 𝑓 † ℓ!1 D𝑥ℓ 𝑓 2,2 if D𝑥 𝑓 is sur-
k𝐴† 𝐵 k 2,2 ≤ 1/(1 − k𝐵 † (𝐴 − 𝐵)k 2,2 ). jective, and zero otheriwse.

Proof of Theorem 5.2. We will apply Theorem 3.1. Note that Again, we can obtain a lower bound of the local reach in terms
† of this new measure of regularity: Smale’s 𝛾. Even though we state
dist2 (𝑧˜ − 𝑧, T𝑧 Z(𝑓 )) = kD𝑧 𝑓 D𝑧 𝑓 [𝑧˜ − 𝑧] k 2, the theorem for polynomials, it holds for more general analylitic
since D𝑧 𝑓 † D𝑧 𝑓 is the orthogonal projection onto (ker D𝑧 𝑓 ) ⊥ = functions. This improves the constant in [4, Theorem 3.3].
(T𝑧 𝑋 ) ⊥ . Hence, assuming kD𝜁 𝑓 † (D𝑧 𝑓 − D𝜁 𝑓 )k 2,2 < 1, Theorem 5.5. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and 𝜁 ∈ Z(𝑓 ). Then
kD𝜁† D𝑧 𝑓 [𝑧˜ − 𝑧] k 2 𝜚 (Z(𝑓 ), 𝜁 ) ≥ 1/(5𝛾 (𝑓 , 𝜁 )).
dist2 (𝑧˜ − 𝑧, T𝑧 Z(𝑓 )) ≤ (5.2)
1− kD𝜁 𝑓 † (D𝑧 𝑓 − D𝜁 𝑓 )k 2,2 Proof. We could repeat the proof of Theorem 5.2 using Theo-
rem 3.1. However, it is easier and equivalent, to deduce this theo-
due to D𝜁 𝑓 being surjective, D𝜁 𝑓 D𝜁 𝑓†
= I and Lemma 5.3. Now,
rem from Theorem 5.2. By expanding D𝑧2 𝑓 in its Taylor series and
by Taylor’s theorem with residue to obtain for some 𝑥 ∈ [𝑧, ˜ 𝑧],
adding up the norms, we have
1
−D𝜁 𝑓 † D𝑧 𝑓 [𝑧˜ − 𝑧] = D𝜁 𝑓 † D𝑥2 𝑓 [𝑧˜ − 𝑧, 𝑧˜ − 𝑧], (5.3) 𝐾 (𝑓 , 𝛾, 𝑟 ) ≤ 2𝛾 (𝑓 , 𝜁 )/(1 − 𝛾 (𝑓 , 𝜁 )𝑟 ) 3 ). (5.7)
2
,, Chris La Valle and Josué Tonelli-Cueto

For 𝑟 = 𝑡/𝛾 (𝑓 , 𝛾), we have 𝐾 (𝑓 , 𝛾, 𝑟 )𝑟 < 2𝑡/(1−𝑡) 3 . Thus 𝐾 (𝑓 , 𝛾, 𝑟 )𝑟 < is 1-Lipschitz with respect the ∞-norm.
𝑡 if 𝑡 < 0.22908 . . .. Hence, by Theorem 5.2, for 𝑡 = 0.22908,
Lemma 6.4. Let 𝐴 ∈ R𝑞×𝑛 be surjective. Then
𝜚 (Z(𝑓 ), 𝜁 ) ≥ 𝑟 = 0.22908/𝛾 (𝑓 , 𝜁 ) ≥ 1/(5𝛾 (𝑓 , 𝜁 )), (5.8)
k𝐴𝑣 k ∞ k𝐴𝑣 k ∞
as desired.  k𝐴† k ∞,2
−1
= min = max min .
𝑣⊥ker 𝐴 k𝑣 k 2 𝑉 ≤R𝑛 𝑣 ∈𝑉 k𝑣 k 2
𝑣≠0 dim 𝑉 =𝑛−𝑞 𝑣≠0
6 CUBIC CONDITION NUMBER AND REACH
Proof of Theorem 6.2. (a) This follows from (c). (b) This is triv-
We introduce a condition number. Instead of following [27], where
ial. (c) By the reverse triangle inequality applied twice and Lemma 6.4,
we only worked in the unit cube, we will be working now in the
|k𝑔k/C(𝑔, 𝑥) − k 𝑓 k 1 /C(𝑓 , 𝑥)| is bounded by the maximum of kΔ −1 k𝑥 k h∞
−𝒅 (𝑔−
whole R𝑛 . Recall that 𝐴† := 𝐴∗ (𝐴𝐴∗ ) −1 is the pseudoinverse of a
𝑓 )(𝑥)k ∞ and
surjective matrix 𝐴 and that, by convention, k𝐴† k ∞,2 = ∞ if 𝐴 is
𝑎 𝑎 𝑎𝑛 − (𝒅 −1)
not surjective. Below, k𝑥 k h∞ := max{1, k𝑥 k ∞ }, 𝜆 := diag(𝜆 1 , . . . , 𝜆 ) min kΔ −2 k𝑥 k h∞ D𝑥 (𝑔 − 𝑓 )[𝑣] k 2 /(k𝑣 k ∞ )}.
𝑛
for 𝜆 > 0 and 𝑎 ∈ R , and 1 := (1, . . . , 1). 𝑣≠0, 𝑣⊥ker 𝐴

Definition 6.1. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and 𝑥 ∈ R𝑛 . Then the local 1- Now, by Corollary 4.3, these both are bounded by kΔ −1 𝑓 k 1 and we
are done. (d) This follows from Proposition 6.3 and
   
condition number is
k 𝑓 k1 1 1 1 1 k𝑦 − 𝑥 k ∞
C(𝑓 , 𝑥) := , − ≤2 .
max{kΔ −1 k𝑥 k −𝒅 𝑓 (𝑥)k ∞ , kD𝑥 𝑓 † k𝑥 k 𝒅−1 Δ2 k ∞,2
h∞
−1 }
h∞
k𝑦 k h∞ 𝑦 k𝑥 k h∞ 𝑥 ∞ k𝑥 k h∞
and, for 𝑅 ∈ [1, ∞], the (global) 1-condition number is (e) We have that
C𝑅 (𝑓 ) := sup C(𝑓 , 𝑧) ∈ [1, ∞]. 1 ℓ 𝒅 −1 Δ2 1 ℓ
D𝑥 𝑓 † D 𝑓 ≤ D𝑥 𝑓 † k𝑥 k h∞ k𝑥 k 1−𝒅
h∞ Δ−2 ℓ! D𝑥 𝑓
k𝑧 k ∞ ≤𝑅 ℓ! 𝑥 2,2 ∞,2 2,∞
6.1 Properties of the Condition Number and, by k𝐴k 2,∞ ≤ k𝐴k ∞,∞ and Corollary 4.3, the second factor is
Theorem 6.2. Let 𝑓 , 𝑔 ∈ P𝑛,𝑑 [𝑞] and 𝑥, 𝑦 ∈ R𝑛 . Then the follow- bounded by
ing hold: 1 − (ℓ −1) ℓ −2
1−𝒅
k𝑥 k h∞ Δ−2 D𝑥ℓ 𝑓 ≤ k𝑥 k h∞ D k 𝑓 k1.
(a) Bounds: 1 ≤ C(𝑓 , 𝑥) ≤ C(𝑓 ) ≤ ∞. ℓ! ∞,∞
(b) Regularity Inequality: Either
Therefore, by (a), (b) and the above,
kΔ −1 k𝑥 k h∞
−𝒅
𝑓 (𝑥)k ∞ /k 𝑓 k 1 > 1/C(𝑓 , 𝑥)
1 ℓ −2
or 𝛾 (𝑓 , 𝑥) ≤ sup C(𝑓 , 𝑥) ℓ −1 D ℓ −1 /k𝑥 k h∞ ≤ C(𝑓 , 𝑥)D/k𝑥 k h∞ ,
2 −1 ℓ ≥2
kD𝑥 𝑓 † k𝑥 k 𝒅−1
h∞ Δ k ∞,2 /k 𝑓 k 1 > 1/C(𝑓 , 𝑥).
which is the claim we wanted to prove. 
(c) 1st Lipschitz Property:
k𝑔k 1 /C(𝑔, 𝑥) − k𝑓 k 1 /C(𝑓 , 𝑥) ≤ kΔ −1 (𝑔 − 𝑓 )k 1 Proof of Proposition 6.3. The equality is an easy consequence
of the definition of homogeneization. For the second part, consider
and
for 𝑥 ∈ 𝐼 1+𝑛 ,
k𝑔k 1 /C𝑅 (𝑔) − k 𝑓 k 1 /C𝑅 (𝑓 ) ≤ kΔ −1 (𝑔 − 𝑓 )k 1 .
 .
k 𝑓 k1
Ch (𝑓 , 𝑥) :=

† 2 −1
(6.1)
(d) 2nd Lipschitz Property:
max Δ −1 𝑓 h (𝑥) ∞, D𝑥 𝑓 h𝑃 0 Δ
|1/C(𝑓 , 𝑦) − 1/C(𝑓 , 𝑥)| ≤ 2k𝑦 − 𝑥 k ∞ /k𝑥 k ∞ ∞,2

kΔ −1 k𝑥 k h∞
−𝒅 𝑓 (𝑥 ) k ∞ We only have to show that
(e) Higher Derivative Estimate: If C(𝑓 , 𝑥) k 𝑓 k1 < 1,
then 𝐼 1+𝑛 ∋ 𝑥 ↦→ 1/Ch (𝑓 , 𝑥)
𝛾 (𝑓 , 𝑥) ≤ D C(𝑓 , 𝑥)/k𝑥 k h∞ . is 1-Lipschitz with respect the ∞-norm.
Proposition 6.3. Let 𝑓 ∈ P𝑛,𝑑 [𝑞]. Denote by 𝑓h ∈ P1+𝑛,𝒅 [𝑞] Assume without loss of generality that k 𝑓 k 1 = 1. By the reverse
the homogeneization of 𝑓 given by triangle inequality and by Lemma 6.4, we only have to show that

𝑓𝑖h := 𝑓𝑖 (𝑋 /𝑋 0)𝑋 0𝑑𝑖 . 𝐼 1+𝑛 ∋ 𝑥 ↦→ Δ −1 𝑓 h (𝑥)

∞
Then, for all 𝑥 ∈ R𝑛 , and
(
k 𝑓 k1
)  † −1
kΔ −2 D𝑥 𝑓 [𝑣] k ∞
   †
C(𝑓 , 𝑥) =
−1 𝐼 1+𝑛 ∋ 𝑥 ↦→ D𝑥 𝑓 h 𝑃0 Δ2 = min
max Δ −1 𝑓 h k𝑥𝑥k , D 𝑥 𝑓 h 𝑃0 Δ2 ∞,2 𝑣≠0 k𝑣 k 2
h∞ ∞ k𝑥 k h∞ 𝑣⊥ker 𝐴
∞,2
are 1-Lipschitz with respect the ∞-norm. Now, by the reverse tri-
where 𝑃0 = I − 𝑒 0𝑒 0∗ . In particular,
  angle inequality again,
𝑥 1
𝜕𝐼 𝑛+1 ∋ 0 ↦→ Δ −1 𝑓 h (𝑦) − Δ −1 𝑓 h (𝑥) ≤ Δ −1 𝑓 h (𝑦) − Δ −1 𝑓 h (𝑥)
𝑥 C(𝑓 , 𝑥/𝑥 0 ) ∞ ∞ ∞
Some Lower Bounds on the Reach of an Algebraic Variety ,,

and Corollary 6.7. Let 𝑓 ∈ P𝑛,𝑑 [𝑞]. Then

𝜚 𝑅 (Z(𝑓 )) ≥ 1/max {D − 2, C𝑅 (𝑓 )} .
kΔ −2 D𝑥 𝑓 [𝑣] k ∞ kΔ −2 D𝑥 𝑓 [𝑣] k ∞
min − min
𝑣≠0 k𝑣 k 2 𝑣≠0 k𝑣 k 2 Remark 6.8. The most interesting aspect of the bound is that the
𝑣⊥ker 𝐴 𝑣⊥ker 𝐴 local reach of points “near to finity” grows linearly with the dis-
kΔ −2 D 𝑦𝑓 [𝑣] − Δ −2 D𝑥 𝑓 [𝑣] k ∞ tance to the origin. The following two examples show what hap-
≤ min
𝑣≠0 k𝑣 k 2 pens when the introduced condition number is not finite.
𝑣⊥ker 𝐴
Example 6.9 (Two parallel lines). The set formed by two parallel
Now, by Corollary 4.4, Δ −1 𝑓 and Δ −2 𝜕𝑓 [𝑣]/k𝑣 k ∞ are 1-Lipschitz
lines, e.g., 𝑓 = 𝑋 (𝑋 − 1), is projectively singular, since the lines
with respect the ∞-norm. Therefore the proof is finished. 
“intersect at infinity”. Thus lim𝑅→∞ C𝑅 (𝑓 ) = ∞. We can see that
Proof of Lemma 6.4. If 𝐴 is not surjective, then the claim fol- 𝜚 (Z(𝑓 ), (𝑥, 𝑦)) = 1, which clearly does not grow linearly at all.
lows since all terms are zero. Now, assume 𝐴 is surjective. We first
Example 6.10. The parabola given by 𝑓 = 𝑌 − 𝑋 2 intersects tan-
prove the first equality and then the second one.
gentially the line at infinity, and so lim𝑅→∞ C𝑅 (𝑓 ) = ∞. Since
We have that
  −1
𝜚 (Z(𝑓 ), (𝑥, 𝑦)) = min{1/2, |𝑥 |},
k𝐴†𝑤 k 2 k𝑤 k ∞ 𝜚 (𝑓 , (𝑥, 𝑦))
k𝐴† k ∞,2
−1
= max = min † . lim = 0.
𝑤≠0 k𝑤 k ∞ 𝑤≠0 k𝐴 𝑤 k 2 𝑥→∞ k (𝑥, 𝑦)k ∞
Now, all values of 𝑤 ≠ 0 are of the form 𝐴𝑣 with 𝑣 ≠ 0 and 𝑣 ⊥ Differently to the case of the two lines, the reach of a parabola
ker 𝐴. Hence grows linearly as we go to infinity, although it does not grow lin-
k𝐴k ∞ k𝐴𝑣 k ∞ k𝐴𝑣 k ∞ early with respect the distance to the origin.
min † = min = min ,
𝑤≠0 k𝐴 𝑤 k 2 𝑣≠0 k𝐴†𝐴𝑣 k 2 𝑣≠0 k𝑣 k 2
𝑣⊥ker 𝐴 𝑣⊥ker 𝐴 Proof of Theorem 6.6. We will apply Theorem 5.2. Using sub-
since 𝐴† 𝐴
is the orthogonal projection onto the orthogonal com- multiplicativity of the norm, we can easily see that 𝐾 (𝑓 , 𝜁 , 𝑟 ) is
plement of ker 𝐴. This shows the first equality. bounded by
For the second equality, we only have to show that for a linear k𝐷𝜁 𝑓 † k𝜁 k 𝒅−1 2 −𝒅+1 −2 2
h∞ Δ k ∞,2 max k𝜁 k h∞ Δ D𝑧 𝑓 .
subsapce 𝑉 ≤ R𝑛 of dimension 𝑛 − 𝑞, min𝑣 ∈𝑉 , 𝑣≠0 k𝐴𝑣 k ∞ /k𝑣 k 2 ≤ 𝑧 ∈𝐵 (𝜁 ,𝑟 ) 2,∞
min𝑣⊥ker 𝐴, 𝑣≠0 k𝐴𝑣 k ∞ /k𝑣 k 2 . If 𝑉 ∩ ker 𝐴 ≠ 0, then this is trivial. By the definition of the condition number and 𝜁 ∈ Z(𝑓 ), we only
Otherwise, the orthogonal projection onto the orthogonal comple- have to focus on the right-hand side factor. Now,
ment of ker 𝐴, 𝐴†𝐴, does not vanish on 𝑉 , and so for all non-zero
−𝒅+1 −2 2
𝑣 ∈ 𝑉, max k𝜁 k h∞ Δ D𝑧 𝑓
𝑧 ∈𝐵 (𝜁 ,𝑟 ) 2,∞
k𝐴𝑣 k ∞ k𝐴𝐴†𝐴𝑣 k ∞ k𝐴†𝐴𝑣 k 2 k𝐴(𝐴†𝐴𝑣)k ∞ −𝒅+1 Δ −2 D2 𝑓
= ≤ . ≤ max k𝜁 k h∞ 𝑧
k𝑣 k 2 k𝐴†𝐴𝑣 k 2 k𝑣 k 2 k𝐴†𝐴𝑣 k 2 𝑧 ∈𝐵 (𝜁 ,𝑟 ) ∞,∞
 
From here the desired inequality follows.  −1 max k𝑧 k h∞ D−2 − (𝒅−21) −2 2
≤ k𝜁 k h∞ k𝑧 k h∞ Δ D𝑧 𝑓
𝑧 ∈𝐵 (𝜁 ,𝑟 ) k𝜁 k h∞ 2,∞
  D−2
We can show also a Condition Number Theorem, which gives a
geometric interpretation of the condition number. We state it with- −1 max k𝑧 k h∞
out a proof, since it is very similar to that of [27, Propositions 4.4 ≤ k 𝑓 k 1 k𝜁 k h∞
𝑧 ∈𝐵 (𝜁 ,𝑟 ) k𝜁 k h∞
& 4.6].
where in the first line we use kIk 2,∞ = 1, in the second line we
Theorem 6.5. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and multiply and divide by k𝑧 k 𝒅−21 and apply submultiplicativity, and
 in the third line we use Corollary 4.3.
Σ𝑅𝑛,𝒅 [𝑞] := 𝑔 ∈ P𝑛,𝑑 [𝑞] | ∃𝜁 ∈ Z ( 𝑔) : k𝜁 k ∞ ≤ 𝑅 .
Hence, under the assumption 𝑟 /k𝜁 k h∞ < 1/(D − 2),
Then −1 C(𝑓 , 𝜁 ),
𝐾 (𝑓 , 𝜁 , 𝑟 ) ≤ k𝜁 k h∞ (6.2)
k 𝑓 k1 (1 + D)k 𝑓 k 1
≤ C(𝑓 ) ≤ . since then k𝑧 k h∞ ≤ 1 + k𝜁 k h∞ /(D − 2). Now, for
dist1 (Δ −1 𝑓 , Σ𝑅𝑛,𝒅 [𝑞]) dist1 (Δ −1 𝑓 , Σ𝑅𝑛,𝒅 [𝑞])  
k𝜁 k h∞ |𝜁 k h∞
 𝑟 = min , ,
D − 2 C(𝑓 , 𝜁 )
6.2 Condition-based Bound on Reach we have 𝐾 (𝑓 , 𝜁 , 𝑟 )𝑟 < 1, and so we can apply Theorem 5.2 to finish
The higher derivative estimate induces a bound on the reach. How- the proof. 
ever, this bound depends on both the condition number and the
degree. The following bound does only depend on the condition
7 RANDOM BOUNDS FOR THE REACH:
number. THE CONTINUOUS CASE
We establish bounds for the reach of the zero set of a zintzo random
Theorem 6.6. Let 𝑓 ∈ P𝑛,𝑑 [𝑞] and 𝜁 ∈ Z(𝑓 ). Then
polynomial. To do this, we do this through the condition-based
𝜚 (Z(𝑓 ), 𝜁 ) ≥ k𝜁 k h∞ /max {D − 2, C(𝑓 , 𝜁 )} . bound of the reach. The techniques in this section are essential for
,, Chris La Valle and Josué Tonelli-Cueto

the next section, where we deal with integer bit random polynomi- (u) If the coefficients of 𝔣 are independent random variables uni-
als which are more technical to handle. formly distributed in intervals of length at least 𝜆 contained in
[−𝑅, 𝑅], then for all 𝑝 ≥ 1, 𝔣 is a 𝑝-zintzo random polynomial
7.1 Zintzo random polynomial tuples tuple such that for all 𝑖, 𝐿𝔣,𝑖 ≤ max𝑖 |𝑀𝑖 |𝑅 and 𝜌 𝔣,𝑖 ≤ 𝜆 −1 .
Zintzo random polynomials where introduced in [26] (cf. [27]) mod- In particular, if all coefficients are uniformly distributed in
elling a robust probabilistic model of polynomials. Here, we extend [−1, 1], then 𝐿𝔣,𝑖 ≤ max𝑖 |𝑀𝑖 | and 𝜌 𝔣,𝑖 = 1/2.
the notion to polynomial tuples and extend the results there to our 
condition number. We recall the following definitions needed for
the definition of zintzo random polynomials. It is important to note that, arguing as in [27, Proposition 7.11],
for a 𝑝-zintzo random polynomial,
Definition 7.1. Let 𝔵 ∈ R be a random variable and 𝑝 ≥ 1. Then:  
(1) 𝔵 has anticoncentration if there is some 𝜌 > 0 such that for 1 1 𝑛+1 𝑛2
𝐿𝔣 𝜌 𝔣 ≥ max |𝑀𝑖 | ≥ ≥ . (7.4)
all 𝜀, 𝑥 ∈ R, P(|𝔵 − 𝑥 | < 𝜀) < 2𝜌𝜀. The smallest such 𝜌 is 4 𝑖 2 2 4
called the anticoncentration constant of 𝔵.
(2) 𝔵 is 𝑝-subexponential (subgaussian if 𝑝 ≥ 0) if there is some 7.2 Condition of a zintzo random polynomial
𝑝 𝑝
𝐿 > 0 such that for all 𝑡 ≥ 𝐿, P(|𝔵| ≥ 𝑡) ≤ e−𝑡 /𝐿 . The The following theorems are standard.
smallest such constant is called the 𝑝-subexponential con-
Theorem 7.4. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a 𝑝-zintzo random polynomial
stant of 𝔵.
tuple with supports 𝑀1 , . . . , 𝑀𝑞 . Then for 𝑡 ≥ e𝑛+1 ,
Now, the definition of zintzo random polynomials adapted to 𝑛+𝑞
our setting. Note that support condition is similar to that of [24], 𝑝+2 𝑞 ln 𝑝 𝑡
P(C𝑅 (𝔣) ≥ 𝑡) ≤ 8𝑅(𝑛 + 1) 2𝑝 𝑛+ 𝑝 +1 D𝑞+2𝑛 (4𝐿𝔣 𝜌 𝔣 )𝑛+𝑞
which is possibly due to the chosen projectivization of R𝑛 . 𝑡
Definition 7.2. Let 𝑝 ≥ 1 and 𝑀1 , . . . , 𝑀𝑞 ⊆ N𝑛 finite set such Theorem 7.5. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a 𝑝-zintzo random polynomial
that for each 𝑖, tuple with supports 𝑀1 , . . . , 𝑀𝑞 and 𝑥 ∈ R𝑛 . Then for 𝑡 ≥ e𝑛+1 ,
Ø
𝑛
𝑛+𝑞
𝑅𝑛,𝑑𝑖 := {(𝑑𝑖 − 1)𝑒𝑘 + 𝑒𝑙 | 𝑙 ∈ {0, . . . , 𝑛}} ∈ 𝑀𝑖 (7.1) 𝑝+2 𝑞 ln 𝑝 𝑡
P(C(𝔣, 𝑥) ≥ 𝑡) ≤ 4k𝑥 k h∞ (𝑛 + 1) 2𝑝 𝑛+ 𝑝 D𝑞+2𝑛 (4𝐿𝔣 𝜌 𝔣 )𝑛+𝑞
𝑘=0
𝑡 𝑛+1
where by convention 𝑒 0 ∈ Z𝑛 is the zero vector. A 𝑝-zintzo random
polynomial 𝑞-tuple supported on 𝑀1 , . . . , 𝑀𝑞 is a random polyno- Proof of Theorem 7.4. Using Proposition 6.3, we will show that
mial tuple 𝔣 ∈ P𝑛,𝑑 [𝑞] with  
1 𝑛
Õ 𝑛
P(C𝑅 (𝔣) ≥ 𝑡) ≤ (𝑛 +1)𝑡 1 + sup P(C(𝔣, 𝑥) ≥ 𝑡/2). (7.5)
𝔣𝑖 = 𝔣𝑖,𝛼 𝑋 𝛼 (𝑖 ∈ {1, . . . , 𝑞}) 2𝑡 k𝑥 k ∞ ≤𝑅
𝛼 ∈𝑀𝑖
Once this is shown, Theorem 7.5 finishes the proof.
such that the coefficients 𝔣𝛼 are independent 𝑝-subexpoential ran- Let G𝑡 ⊂ 𝐶 := R𝑛+1 𝑛+1 ∩ {𝑥 | 𝑥 ≥ 1/𝑅} be a finite subset
>0 ∩ 𝜕𝐼 0
dom variables with the anti-concentration property. such that for all 𝑥 ∈ 𝐶, dist∞ (𝑐, G𝑡 ) ≤ 1/𝑡. Now, by Proposition 6.3,
Given a 𝑝-zintzo random polynomial tuple 𝔣 ∈ P𝑛,𝑑 [𝑞], we de-
fine the following quantities: C(𝔣) = sup C(𝑓 , 𝑥/𝑥 0 ),
(𝑥 0 ,𝑥 ) ∈𝑋
(1) The tail constants of 𝔣 given by
Õ and, since (𝑥 0 , 𝑥) ↦→ 1/C(𝑓 , 𝑥/𝑥 0 ) is 1-Lipschitz, we have that
𝐿𝔣 := max 𝐿𝑖,𝛼 , (7.2)
𝑖 𝛼 ∈𝑀𝑖
C(𝔣) ≤ 2 sup C(𝔣, 𝑔/𝑔0 )
where 𝐿𝑖,𝛼 is the 𝑝-subexponential constant of 𝔣𝑖,𝛼 . (𝑔0,𝑔) ∈ G𝑡
(2) The anti-concentration constant of 𝔣 given by
whenever 2 C(𝔣)𝑡 ≤ 1. Thus C(𝔣) ≥ 𝑡 implies
𝜌 𝔣 := max 𝜌𝑖,(𝑑𝑖 −1)𝑒𝑘 +𝑒𝑙 (7.3)
𝑘,𝑙 sup C(𝔣, 𝑔/𝑔0 ) > 𝑡/2,
where 𝜌𝑖,𝛼 is the anti-concentration constant of 𝔣𝑖,𝛼 . (𝑔0,𝑔) ∈ G𝑡

It is important to note that Kac polynomial tuples and polynomi- and so, by the implication bound,
als with noise given by Kac random polynomials are zintzo random !
polynomial tuples. P(C(𝔣) ≥ 𝑡) ≤ P sup C(𝔣, 𝑔/𝑔0 ) ≥ 𝑡/2 ,
(𝑔 ,𝑔) ∈ G
0 𝑡
Proposition 7.3. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a random polynomial tuple
with supports 𝑀1 , . . . , 𝑀𝑞 satisfying (7.1). The following holds: and, by the union bound,
(g) If the coefficients of 𝔣 are independent normal variables of P(C(𝔣) ≥ 𝑡) ≤ |G𝑡 | sup P(C(𝔣, 𝑥/𝑥 0 ) ≥ 𝑡/2).
mean contained in [−𝑅, 𝑅] and variance at least 𝜎02 and at (𝑥 0 ,𝑥 ) ∈𝑋
most 𝜎12 , then 𝐿𝔣,𝑖 = max{𝑅, 2𝜎1 } max𝑖 |𝑀𝑖 | and 𝜌 𝔣,𝑖 = (2𝜋) −1/2 𝜎0−1 .  𝑛
Moreover, if all coefficients are standard normal variables, then Finally, we can achieve |G𝑡 | ≤ (𝑛 + 1)𝑡 𝑛 1 + 2𝑡1 by putting the
𝔣 is a√2-zintzo random polynomial tuple such that for all 𝑖, points in a uniform cubical grid, and sup (𝑥 0,𝑥 ) ∈𝑋 P(C(𝔣, 𝑥/𝑥 0 ) ≥
𝐿𝔣 = 2 max𝑖 |𝑀𝑖 | and 𝜌 𝔣,𝑖 = (2𝜋) −1/2 . 𝑡/2) = sup k𝑥 k ∞ ≤𝑅 P(C(𝔣, 𝑥) ≥ 𝑡/2). Thus (7.5) follows. 
Some Lower Bounds on the Reach of an Algebraic Variety ,,

Proof of Theorem 7.5. Consider Now, to apply the above result, we consider the linear projec-
!
tion,
n o.
k 𝑓 k1
C(𝑓 , 𝑥, 𝑣) :=
max −𝒅 𝑓 (𝑥)
Δ −1 k𝑥 k h∞ 1−𝒅 Δ −2 D 𝑓 k /√𝑛
, k𝑣 ∗ k𝑥 k h∞
−𝒅 Δ −1 𝑓 (𝑥)
k𝑥 k h∞
∞
𝑥 2 𝑓 ↦→ √ ∈ R𝑞+𝑛 (7.9)
(k𝑥 k 1−𝒅
h∞
Δ −2 D𝑥 𝑓 ) ∗ 𝑣/ 𝑛
We can easily see, using Lemma 6.4, that
with 𝑈 = 𝑢/𝑡𝐼 𝑞 × B𝑞 , where B𝑞 is the 𝑞-ball. Stirling’s approxima-
C(𝑓 , 𝑥) ≤ max C(𝑓 , 𝑥, 𝑣) (7.6)
𝑣 ∈S𝑞 −1 tion [3, (2.14)] shows that
and, using standard norm inequalities and Corollary 4.3 as in the 1 2𝑛 (e𝜋)𝑞/2𝑢𝑞+𝑛
proof of the 2nd Lipschitz property of Theorem 6.2, that the map vol 𝑈 ≤ √ .
2𝜋 𝑞𝑞/2𝑡 𝑞+𝑛
S𝑞−1 ∋ 𝑣 ↦→ 1/C(𝑓 , 𝑥, 𝑣) (7.7)
If 1 = k𝑥 k h∞ , then obtain following submatrix of (7.9)
is 1-Lipschitz with respect both k k 2 and the geodesic distance.
Δ −1 𝑥 1 Δ −1 ··· 𝑥𝑛 Δ −1
Now, let 𝔳∗ ∈ S𝑞 , be the random vector such that © √ ª
­ 𝑣 ∗ Δ −2 / 𝑛 ®
­ ®
­ ®
max C(𝔣, 𝑥, 𝑣) = C(𝔣, 𝑥, 𝔳∗ ). (7.10)
­ ®
𝑣 ∈S𝑞 −1 ..
.
√
Then, if max𝑣 ∈S𝑞 −1 C(𝔣, 𝑥, 𝑣) ≥ 𝑡, we have have that for all 𝑣 ∈ « 𝑣 ∗ Δ −2 / 𝑛 ¬
𝐵 S (𝔳∗, 1/𝑡) ⊂ S𝑞−1 , C(𝔣, 𝑥, 𝑣) ≥ 𝑡/2. Hence max𝑣 ∈S𝑞 −1 C(𝔣, 𝑥, 𝑣) ≥ 𝑡
implies that by looking at the monomials 1, 𝑋 1, . . . , 𝑋𝑛 . In the above matrix,
each column block corresponds to one of the monomials, corre-
vol 𝐵 S (𝔳∗, 1/𝑡) 1 sponding the columns in each block to the 𝑞 polynomials. If |𝑥𝑖 | =
P𝔳∈S𝑞 −1 (C(𝔣, 𝑥, 𝔳) ≥ 𝑡/2) ≥ ≥ √
vol S𝑞 2𝜋𝑞𝑡 𝑞−1 k𝑥 k h∞ , then we obtain, after adding to the second row multiples
of rows of the first row block,
where the last inequality follows from [3, Lemma 2.34]. Thus
Δ −1 Δ −1 |𝑥𝑖 | −1 𝑥 1 |𝑥𝑖 | −1 Δ −1 𝑥𝑛 |𝑥𝑖 | −1 Δ −1
© √ ª
···
√ √
P𝔣 (C(𝑓 , 𝑥) ≥ 𝑡) ­ −|𝑥𝑖 | −1 𝑣 ∗ Δ −2 / 𝑛 −|𝑥𝑖 | −1 𝑥 1 𝑣 ∗ Δ −2 / 𝑛 −|𝑥𝑖 | −1 𝑥𝑛 𝑣 ∗ Δ −2 / 𝑛 ®
­ ®
···
­ ®
√
𝑣 ∗ Δ −2 / 𝑛
­ ®
≤ P( max C(𝑓 , 𝑥, 𝑣) ≥ 𝑡) ­ ® (7.11)
­ ®
­ ®
 
.
𝑣 ∈S𝑞 −1 ­
.
. ®
√
1 « 𝑣 ∗ Δ −2 / 𝑛 ¬
≤ P𝔣 P𝔳∈S𝑞 −1 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2) ≥ √
2𝜋𝑞𝑡 𝑞−1
p 𝑞−1
(7.8) looking at the monomials𝑋𝑘𝑑𝑖 , 𝑋𝑘𝑑𝑖 −1 , 𝑋𝑘𝑑𝑖 −1 𝑋 1, . . . , 𝑋𝑘𝑑𝑖 −1 𝑋𝑛 , where
≤ 2𝜋𝑞𝑡 E𝔣 P𝔳∈S𝑞 −1 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2) we omit 𝑋𝑘𝑑𝑖 −1 𝑋𝑘 in the latter list. Hence, in both cases, after an
p 𝑞−1
≤ 2𝜋𝑞𝑡 E𝔳∈S𝑞 −1 P𝔣 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2) easy computation and applying the Cauchy-Binet formula, we get
p 𝑞−1 √ √
≤ 2𝜋𝑞𝑡 sup P𝔣 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2) −1 − 𝑛2 −𝑞−2𝑛
det 𝐴𝐴∗ ≥ 𝑞k𝑥 k h∞ 𝑛 D . (7.12)
𝑣 ∈S𝑞 −1
where in the first two inequalities we use the implication bound; Therefore P(C(𝔣, 𝑥, 𝑣) ≥ 𝑡/2) is upper bounded by
!
Õ
in the third one, Markov’s Inequality; and in the fourth one, the 𝑛 𝑞/2 𝑛 𝑞+2𝑛 𝑢 𝑞+𝑛 𝜌 𝑛+𝑞
Fubini-Tonelli theorem.
𝑝
−𝑢 𝑝 /𝐿𝔣 k𝑥 k h∞ 2 (e𝜋) 𝑛 2 D 𝔣
|𝑀𝑖 | e + √ ,
All this reduces us to estimate 2𝜋 𝑞 (𝑞+1)/2𝑡 𝑞+𝑛
𝑖
P𝔣 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2). and so P(C(𝔣, 𝑥) ≥ 𝑡) by
!
Õ
By the union bound, the implication bound and [27, Proposition
p
𝑛 𝑛+𝑞
𝑝
−𝑢 𝑝 /𝐿𝔣
2𝑛 (e𝜋)𝑞/2 𝑛 2 D𝑞+2𝑛 𝑢𝑞+𝑛 𝜌 𝔣
7.4], we have that for 𝑢 ≥ 𝐿𝔣 , 2𝜋𝑞 |𝑀𝑖 | e + k𝑥 k h∞ .
! 𝑞 (𝑞−1)/2𝑡 1+𝑛
Õ 𝑝
−𝑢 𝑝 /𝐿𝔣
𝑖
P𝔣 (C(𝑓 , 𝑥, 𝑣) ≥ 𝑡/2) ≤ |𝑀𝑖 | e Taking 𝑢 = 𝐿𝔣 ((𝑛 + 1) ln 𝑡) 1/𝑝 and estimating gives the bound. We
 
𝑖
√ use that |𝑀𝑖 | ≤ (2D)𝑛 among others. 
−𝒅 −1
+P𝔣 max{k k𝑥 k h∞ Δ 𝑓 (𝑥)k ∞ , k𝑣 ∗ k𝑥 k 1−𝒅 −2
h∞ Δ D𝑥 𝑓 k 2 / 𝑛} ≤ 𝑢/𝑡 ,
where the implication used is that either the numerator of C(𝑓 , 𝑥, 𝑣)
7.3 Bounds on Reach: Proof of Theorem 2.3
is ≥ 𝑢 or the denominator ≤ 𝑢/𝑡. Theorem 2.3 is a particular case of the following theorem.
To bound the last summand, we will use [27, Proposition 7.7]. Theorem 7.6. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a 𝑝-zintzo random polynomial
This proposition, a explicit refinement of [25, Theorem 1.1], states
tuple with supports 𝑀1 , . . . , 𝑀𝑞 . Then for 𝜀 ∈ (0, min{1/D, 1/e− (𝑛+1) }],
that given a random vector 𝔵 ∈ R𝑁 with independent coordinates
P(𝜚 𝑅 (Z(𝑓 )) ≤ 𝜀) is at most
 
with anticoncentration constants ≤ 𝜌, a surjective matrix 𝐴 ∈
R𝑘 ×𝑁 and 𝑈 ⊆ R𝑘 , then 𝑝+2 𝑞 𝑛+𝑞 1
p
𝑛+ +1
√ 8𝑅(𝑛 + 1) 2𝑝 𝑝 D𝑞+2𝑛 (4𝐿𝔣 𝜌 𝔣 )𝑛+𝑞 𝜀 ln 𝑝
P(𝐴𝔵 ∈ 𝑈 ) ≤ vol(𝑈 )( 2𝜌)𝑘 / det(𝐴𝐴∗ ). 𝜀

p to find appropiate sub-

By the Cauchy-Binet formula, we only have Proof of Theorem 7.6. This is Corollary 6.7 combined with The-
matrices of 𝐴 to obtain lower bounds on det(𝐴𝐴∗ ). orem 7.4. 
,, Chris La Valle and Josué Tonelli-Cueto

8 INTEGER BOUNDS FOR THE REACH Theorem 8.6. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be be a random bit polynomial
8.1 Worst-case: Proof of Theorem 2.1 tuple of bit-size 𝜏 and 𝑥 ∈ R𝑛 . Then for 𝑡 ≤ 𝑞(𝑛 + 1) max𝑖 |𝑀𝑖 | 2𝜏 −1 ,
  𝑞+𝑛
Proof of Theorem 2.1. By (6.1) and arguing as in the proof of 𝑛 √ 1
Theorem 7.4, we can see that C𝑅 (𝑓 ) is bounded by P(C(𝔣, 𝑥) ≥ 𝑡) ≤ 6k𝑥 k h∞𝑛 2 2 max |𝑀𝑖 | D𝑞+2𝑛 e𝑢 (𝔣) 𝑛+1 .
𝑖 𝑡
D2 k 𝑓 k 1 / min max{k 𝑓 h (𝑧)k ∞, k𝑣 ∗ D𝑧 𝑓 h 𝑃0 k 2 } Remark 8.7. Note that Theorems 8.5 and 8.6 are meaningful only
𝑧, 𝑣 ∈S𝑞 −1
for 𝜏 large enough.
where 𝑧 ∈ 𝜕𝐼 𝑛+1 satisfies that for each 𝑖, −𝑅𝑧 0 ≤ 𝑧𝑖 ≤ 𝑅𝑧 0 . Now, by
assumption, D2 k 𝑓 k 1 ≤ D2 2𝜏 . Hence, we only need to handle the
Proof of Theorem 8.5. This is (7.5) with Theorem 8.6. 
denominator. For this, we will use [21, Theorem 1] by Jeronimo,
Perrucci and Tsigaridas.
Now, consider the semialgebraic set 𝑆 ⊆ R1+𝑞+2𝑛 given by Proof of Theorem 8.6. The proof follows the same steps as that
of Theorem 8.6. However, it differs in two points: (1) There is no
±𝑧𝑖 ≤ 𝑅𝑧 0, k𝑣 k 2 = 1 𝑡 ≥ ±𝑓𝑖h (𝑥), 𝑡 2 ≥ k𝑤 k 22, 𝑤 = D𝑧 𝑓 h need to control the norm, since k 𝑓 k 1 ≤ max |𝑀𝑖 |2𝜏 . Hence
where (𝑥, 𝑣, 𝑤, 𝑡) ∈ R𝑛 × R𝑞 × R𝑛 × R. This semialgebraic set is
described by 2+4𝑛 polynomials of degree at most D and coefficients P𝔣 (C(𝔣, 𝑥, 𝑣)
of bit size at most max{log 𝑅, 𝜏 }. Moreover,  
−𝒅 −1
≤ P𝔣 max{k k𝑥 k h∞ Δ 𝑓 (𝑥)k ∞ , k𝑣 ∗ k𝑥 k 1−𝒅 −2
h∞ Δ D𝑥 𝑓 k 2 } ≤ 𝑢/𝑡
min 𝑡= min max{k 𝑓 h (𝑧)k ∞, k𝑣 ∗ D𝑧 𝑓 h 𝑃0 k 2 }.  
(𝑥,𝑣,𝑤,𝑡 ) ∈𝑆 𝑧, 𝑣 ∈S𝑞 −1 −𝒅 −1
≤ P𝔣 max{k k𝑥 k h∞ Δ 𝑓 (𝑥)k ∞ , k𝑣 ∗ k𝑥 k 1−𝒅 −2
h∞ Δ D𝑥 𝑓 k ∞ } ≤ 𝑢/𝑡
Hence, by [21, Theorem 1], C𝑅 (𝑓 ) = ∞ or
log C𝑅 (𝑓 ) ≤ 4𝑛(2D) 1+𝑞+2𝑛 (5 + 𝜏 + log 𝑅 + 6𝑛 log D) + 2 log D + 𝜏. where 𝑢 = max |𝑀𝑖 | 2𝜏 . (2) Where we applied [27, Proposition 7.7],
we apply now [15, Proposition 2.7] to the map (7.9). This proposi-
Using Theorem 6.6, the desired bound follows.  tion states that for 𝔵 ∈ Z𝑁 random such that all probability weights
Remark 8.1. Unfortunately, the usual formulations of the reach do are ≤ 𝑤, 𝐴 ∈ R𝑘 ×𝑁 and 𝑏 ∈ R𝑘 , we have that for 𝜀 ∈ [k𝐴k ∞,∞, ∞),
not allow for the application of [21, Theorem 1].  √ 𝑘 √
P(k𝐴𝔵 + 𝑏 k ∞ ≤ 𝜀) ≤ 2 2 2𝑤𝜀 / det 𝐴𝐴∗ .
8.2 Random bit polynomials
We extend the definition of random bit polynomials from [15] to Moreover, conditioning on all random coefficients that are not in
the multivariate setting in which we are working. the 𝑅𝑛,𝑑𝑖 (see (7.1)), we can assume without loss of generality that
𝑀𝑖 = 𝑅𝑛,𝑑𝑖 . Then the matrices of the affine map (7.9) are still (7.10)
Definition 8.2. Let 𝑀1 , . . . , 𝑀𝑞 ⊆ N𝑛 be finite sets satisfying (7.1). and (7.11). In both cases, we have k𝐴k ∞,∞ ≤ 𝑞(𝑛 + 1). Hence
Í
A random bit polynomial 𝑞 tuple with bit-size 𝜏 is a random polyno- P(C(𝔣, 𝑥, 𝑣) ≥ 𝑡/2) is upper bounded by
mial tuple 𝔣 ∈ P𝑛,𝑑 [𝑞] with 𝔣𝑖 = 𝛼 ∈𝑀𝑖 𝔣𝑖,𝛼 𝑋 𝛼 for 𝑖 ∈ {1, . . . , 𝑞},   𝑞+𝑛
such that the coefficients 𝔣𝛼 are independent integers of absolute 1 𝑛 √
2𝑞 − 2 k𝑥 k h∞𝑛 2 D𝑞+2𝑛 2 2𝑤 (𝔣) max |𝑀𝑖 |2𝜏 𝑡 − (𝑞+𝑛)
value at most 2𝜏 . 𝑖
Given such a random polynomial 𝔣, its weight is
 whenever 𝑡 ≤ 𝑞(𝑛 +1) max𝑖 |𝑀𝑖 | 2𝜏 . Now, 2𝑤 (𝔣)2𝜏 ≤ e𝑢 (𝔣) and (7.8)
𝑤 (𝔣) := min P(𝔣𝑖,𝛼 = 𝑐) | 𝑐 ∈ Z, (𝑖, 𝛼) ∈ ∪𝑛𝑖=1 {𝑖} × 𝑅𝑛,𝑑𝑖 ; (8.1)
gives the desired bound. 
and its uniformity is
 
𝑢 (𝔣) := ln (1 + 2𝜏+1 ) 𝑤 (𝔣) . (8.2)
8.4 Probabilistic case: Proof of Theorem 2.2
Theorem 2.2 is a particular case of the following theorem.
Example 8.3. A uniform random bit polynomial 𝑞-tuple of bit-size
𝑡𝑎𝑢 is a random polynomial 𝑞-tuple 𝔣 ∈ P𝑛,𝑑 [𝑞] whose coefficients Theorem 8.8. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a random bit polynomial tuple
are independent uniformly distributed integers in [−2𝜏 , 2𝜏 ]. Such of bit-size 𝜏. Then for 𝜀 ∈ [2/(𝑞(𝑛 + 1) max𝑖 |𝑀𝑖 | 2𝜏 ) , min{1/D, 1/𝑛}],
a polynomial is a random bit polynomial with uniformity 1.   𝑞+𝑛
𝑛
+1
√
Remark 8.4. As shown in [15], this random model has the advan- P(𝜚 𝑅 (Z(𝔣)) ≤ 𝜀) ≤ 20𝑅𝑛 2 2 max |𝑀𝑖 | D𝑞+2𝑛 e𝑢 (𝔣) 𝜀.
𝑖
tage that it is robust under change of support, selection of signs
and selection of exact bit-size among others. Remark 8.9. Note that Theorem 8.8 is meaningful only for 𝜏 large
enough.
8.3 Condition of random bit polynomials
We know prove the discrete analogue of Theorem 7.4 and 7.5. Proof of Theorem 8.8. We just apply Theorems 8.5 and 6.6. 

Theorem 8.5. Let 𝔣 ∈ P𝑛,𝑑 [𝑞] be a random bit polynomial tuple

Acknowledgements. J.T.-C. grateful to Evgenia Lagoda for moral
of bit-size 𝜏. Then for 𝑡 ∈ [𝑛, 𝑞(𝑛 + 1) max𝑖 |𝑀𝑖 | 2𝜏 −2 ],
  𝑞+𝑛
support and to Jazz G. Suchen for useful suggestions for this pa-
𝑛 √ 1 per. C.L.V. and J.T.-C. were partially supported by start-up funds of
P(C𝑅 (𝔣) ≥ 𝑡) ≤ 20𝑅𝑛 2 +1 2 max |𝑀𝑖 | D𝑞+2𝑛 e𝑢 (𝔣) .
𝑖 𝑡 Alperen Ergür and NSF CCF 2110075.
Some Lower Bounds on the Reach of an Algebraic Variety ,,

REFERENCES [24] J. Renegar. 1987. On the efficiency of Newton’s method in approximating all
[1] Eddie Aamari, Jisu Kim, Frédéric Chazal, Bertrand Michel, Alessandro Rinaldo, zeros of a system of complex polynomials. Math. Oper. Res. 12, 1 (1987), 121–148.
and Larry Wasserman. 2019. Estimating the reach of a manifold. Electron. J. Stat. https://doi.org/10.1287/moor.12.1.121
13, 1 (2019), 1359–1399. https://doi.org/10.1214/19-ejs1551 [25] Mark Rudelson and Roman Vershynin. 2015. Small ball probabilities for linear
[2] Paul Breiding, Sara Kališnik, Bernd Sturmfels, and Madeleine Weinstein. 2018. images of high-dimensional distributions. Int. Math. Res. Not. IMRN 19 (2015),
Learning algebraic varieties from samples. Rev. Mat. Complut. 31, 3 (2018), 545– 9594–9617. https://doi.org/10.1093/imrn/rnu243
593. https://doi.org/10.1007/s13163-018-0273-6 [26] Josué Tonelli-Cueto and Elias Tsigaridas. [2020] ©2020. Condition numbers for
[3] Peter Bürgisser and Felipe Cucker. 2013. Condition. Grundlehren der the cube. I: Univariate polynomials and hypersurfaces. In ISSAC’20—Proceedings
mathematischen Wissenschaften [Fundamental Principles of Mathe- of the 45th International Symposium on Symbolic and Algebraic Computation.
matical Sciences], Vol. 349. Springer, Heidelberg. xxxii+554 pages. ACM, New York, 434–441.
https://doi.org/10.1007/978-3-642-38896-5 The geometry of numerical [27] Josué Tonelli-Cueto and Elias Tsigaridas. 2023. Condition numbers for the cube.
algorithms. I: Univariate polynomials and hypersurfaces. J. Symbolic Comput. 115 (2023),
[4] P. Bürgisser, F. Cucker, and P. Lairez. 2019. Computing the homology of basic 142–173. https://doi.org/10.1016/j.jsc.2022.08.013
semialgebraic sets in weak exponential time. J. ACM 66, 1 (2019), Art. 5, 30. [28] R. Vershynin. 2018. High-dimensional probability: An introduction with
https://doi.org/10.1145/3275242 [Publication date initially given as 2018]. applications in data science. Cambridge Series in Statistical and Proba-
[5] P. Bürgisser, F. Cucker, and J. Tonelli-Cueto. 2020. Computing the Homology of bilistic Mathematics, Vol. 47. Cambridge University Press, Cambridge.
Semialgebraic Sets. I: Lax Formulas. Foundations of Computational Mathematics https://doi.org/10.1017/9781108231596
20, 1 (2020), 71–118. https://doi.org/10.1007/s10208-019-09418-y On-line from
May of 2019.
[6] P. Bürgisser, F. Cucker, and J. Tonelli-Cueto. 2021. Computing the Homology
of Semialgebraic Sets. II: General formulas. Foundations of Computational Math-
ematics 21, 5 (1 2021), 1279–1316. https://doi.org/10.1007/s10208-020-09483-8
On-line from January of 2021. arXiv: 1903.10710.
[7] F. Cucker, T. Krick, and M. Shub. 2018. Computing the homology
of real projective sets. Found. Comput. Math. 18, 4 (2018), 929–970.
https://doi.org/10.1007/s10208-017-9358-8
[8] J.-P. Dedieu. 2006. Points fixes, zéros et la méthode de Newton. Mathématiques
& Applications (Berlin) [Mathematics & Applications], Vol. 54. Springer, Berlin.
xii+196 pages.
[9] Sandra Di Rocco, Parker B. Edwards, David Eklund, Oliver Gäfvert, and
Jonathan D. Hauenstein. 2023. Computing geometric feature sizes for al-
gebraic manifolds. SIAM J. Appl. Algebra Geom. 7, 4 (2023), 716–741.
https://doi.org/10.1137/22M1522656
[10] Sandra Di Rocco, David Eklund, and Oliver Gäfvert. 2022. Sampling
and homology via bottlenecks. Math. Comp. 91, 338 (2022), 2969–2995.
https://doi.org/10.1090/mcom/3757
[11] Sandra Di Rocco, David Eklund, and Madeleine Weinstein. 2020. The bottleneck
degree of algebraic varieties. SIAM J. Appl. Algebra Geom. 4, 1 (2020), 227–253.
https://doi.org/10.1137/19M1265776
[12] David Eklund. 2023. The numerical algebraic geometry of bot-
tlenecks. Adv. in Appl. Math. 142 (2023), Paper No. 102416, 20.
https://doi.org/10.1016/j.aam.2022.102416
[13] Ioannis Emiris, Bernard Mourrain, and Elias Tsigaridas. 2020. Separation
bounds for polynomial systems. J. Symbolic Comput. 101 (2020), 128–151.
https://doi.org/10.1016/j.jsc.2019.07.001
[14] Ioannis Z. Emiris, Bernard Mourrain, and Elias P. Tsigaridas. 2010. The DMM
bound: multivariate (aggregate) separation bounds. In ISSAC 2010—Proceedings
of the 2010 International Symposium on Symbolic and Algebraic Computation.
ACM, New York, 243–250. https://doi.org/10.1145/1837934.1837981
[15] Alperen Ergür, Josué Tonelli-Cueto, and Elias Tsigaridas. [2022] ©2022. Beyond
worst-case analysis for root isolation algorithms. In ISSAC ’22—Proceedings of
the 2022 International Symposium on Symbolic and Algebraic Computation. ACM,
New York, 139–148.
[16] Alperen A. Ergür, Grigoris Paouris, and J. Maurice Rojas. 2019. Probabilis-
tic condition number estimates for real polynomial systems I: A broader
family of distributions. Found. Comput. Math. 19, 1 (2019), 131–157.
https://doi.org/10.1007/s10208-018-9380-5
[17] Alperen A. Ergür, Grigoris Paouris, and J. Maurice Rojas. 2021. Smoothed analy-
sis for the condition number of structured real polynomial systems. Math. Comp.
90, 331 (2021), 2161–2184. https://doi.org/10.1090/mcom/3647
[18] H. Federer. 1959. Curvature measures. Trans. Amer. Math. Soc. 93 (1959), 418–491.
https://doi.org/10.2307/1993504
[19] Emil Horobeţ. 2024. The critical curvature degree of an algebraic
variety. J. Symbolic Comput. 121 (2024), Paper No. 102259, 12.
https://doi.org/10.1016/j.jsc.2023.102259
[20] Emil Horobeţ and Madeleine Weinstein. 2019. Offset hypersurfaces and persis-
tent homology of algebraic varieties. Comput. Aided Geom. Design 74 (2019),
101767, 14. https://doi.org/10.1016/j.cagd.2019.101767
[21] Gabriela Jeronimo, Daniel Perrucci, and Elias Tsigaridas. 2013. On the minimum
of a polynomial function on a basic closed semialgebraic set and applications.
SIAM J. Optim. 23, 1 (2013), 241–255. https://doi.org/10.1137/110857751
[22] Partha Niyogi, Stephen Smale, and Shmuel Weinberger. 2008. Finding the homol-
ogy of submanifolds with high confidence from random samples. Discrete Com-
put. Geom. 39, 1-3 (2008), 419–441. https://doi.org/10.1007/s00454-008-9053-2
[23] P. Niyogi, S. Smale, and S. Weinberger. 2011. A topological view of unsu-
pervised learning from noisy data. SIAM J. Comput. 40, 3 (2011), 646–663.
https://doi.org/10.1137/090762932