DOKK Library

On a Generalization of Iterated and Randomized Rounding

Authors Nikhil Bansal,

License CC-BY-3.0

Plaintext
                           T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23
                                        www.theoryofcomputing.org




        On a Generalization of Iterated and
             Randomized Rounding
                                                   Nikhil Bansalโˆ—
               Received April 24, 2019; Revised November 18, 2022; Published December 5, 2024




       Abstract. We give a general method for rounding linear programs, that combines
       the commonly used iterated rounding and randomized rounding techniques. In
       particular, we show that whenever iterated rounding can be applied to a problem with
       some slack, there is a randomized procedure that returns an integral solution that
       satisfies the guarantees of iterated rounding and also has concentration properties.
       We use this to give new results for several classical problems such as rounding
       column-sparse LPs, makespan minimization on unrelated machines and degree-
       bounded spanning trees.


1     Introduction
A powerful approach in approximation algorithms is to formulate the problem at hand as a
0-1 integer program and consider some efficiently solvable relaxation for it. Then, given some
fractional solution ๐‘ฅ โˆˆ [0, 1]๐‘› to this relaxation, apply a suitable rounding procedure to ๐‘ฅ to
     An extended abstract of this paper appeared in the Proceedings of the 51st Annual ACM Symposium on Theory
of Computing (STOCโ€™19) [7].
   โˆ— Supported by the ERC Consolidator Grant 617951 and the NWO VICI grant 639.023.812.



ACM Classification: F.2.2, G.1.6
AMS Classification: 68W25, 68W20
Key words and phrases: approximation, randomized rounding, scheduling, discrepancy,
semidefinite programming, random walks


ยฉ 2024 Nikhil Bansal
c b Licensed under a Creative Commons Attribution License (CC-BY)                  DOI: 10.4086/toc.2024.v020a006
                                            N IKHIL BANSAL

obtain an integral 0-1 solution. Arguably, the two most basic and extensively studied techniques
for rounding such relaxations are randomized rounding and iterated rounding.

Randomized rounding. Here, the fractional values ๐‘ฅ ๐‘– โˆˆ [0, 1] are interpreted as probabilities,
and used to round the variables independently to 0 or 1. A key property of this rounding is that
each linear constraint is preserved in expectation, and its value is tightly concentrated around
its mean as given by Chernoff bounds, or more generally Bernsteinโ€™s inequality. Randomized
rounding is well-suited to problems where the constraints do not have much structure, or when
they are soft and some error can be tolerated. Sometimes these errors can be fixed by applying
problem-specific alteration steps. We refer to [36, 35] for various applications of randomized
rounding.

Iterated rounding. This technique is quite different from randomized rounding and is useful
for problems that may have some hard combinatorial constraints that must be maintained, e. g.,
if the final solution must be a spanning tree or a matching. It is also useful for problems where
the constraints may have some other interesting structural property such as column-sparsity
that we may wish to exploit.
     The rounding is based on linear algebra and it proceeds in several iterations ๐‘˜ = 1, 2, . . .,
until all variables are eventually rounded to 0 or 1. More specifically, we start with ๐‘ฅ (0) = ๐‘ฅ
initially, and let ๐‘ฅ (๐‘˜โˆ’1) โˆˆ โ„ ๐‘› be the solution at the beginning of iteration ๐‘˜ and ๐‘› ๐‘˜ denote the
number of fractional variables in ๐‘ฅ (๐‘˜โˆ’1) (i. e., those strictly between 0 and 1). Then one cleverly
chooses some collection of linear constraints on these ๐‘› ๐‘˜ fractional variables, say specified by
rows of the matrix ๐‘Š (๐‘˜) with rank(๐‘Š (๐‘˜) ) โ‰ค ๐‘› ๐‘˜ โˆ’ 1, and updates the solution as ๐‘ฅ (๐‘˜) = ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜)
by some non-zero vector ๐‘ฆ (๐‘˜) satisfying ๐‘Š (๐‘˜) ๐‘ฆ (๐‘˜) = 0, so that some fractional variable reaches 0
or 1. Such a ๐‘ฆ (๐‘˜) exists as the null space of ๐‘Š (๐‘˜) has dimension ๐‘› ๐‘˜ โˆ’ rank(๐‘Š (๐‘˜) ) โ‰ฅ 1. Notice that
once a variable reaches 0 or 1 it stays fixed.
     Despite its simplicity, this method is extremely powerful and most fundamental results
in combinatorial optimization such as the integrality of matroid, matroid-intersection and
non-bipartite matching polytopes follow very cleanly using this approach. Similarly, several
breakthrough results for problems such as degree-bounded spanning trees, survivable network
design and rounding for column-sparse LPs were obtained by this method. An excellent
reference is [23].

1.1   Need for combining the approaches
In many problem settings, however, one needs a rounding that combines the features of both
randomized and iterated rounding [15, 16, 19, 3]. We give several examples in Section 1.3, but a
typical scenario is where the problem involves finding an object with specific combinatorial
constraints that cannot be violated, e. g., a spanning tree to connect nodes in a network, or
a one-sided matching (assignment) of jobs to machines; and additionally a list of other soft
side-constraints, e. g., a bound on the maximum degree of the spanning tree to prevent any
particular node from being overloaded, or perhaps edges are of several types and we wish to

                       T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                            2
                  O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

purchase a certain minimum number of each type due to fairness considerations, or there may
be multiple budget constraints for various subsets of edges.
   As the soft constraints are typically arbitrary and lack structure, essentially the best one can
hope for is to satisfy them fractionally and then apply randomized rounding. On the other
hand, randomized rounding can be quite bad at satisfying the hard combinatorial constraints,
and iterated rounding is the right approach to handle them. So given a problem with both hard
and soft constraints, either technique by itself does not suffice and one would like a rounding
that simultaneously does as well as iterated rounding on the hard constraints and as well as
randomized rounding on the soft constraints.


Dependent rounding. Motivated by such problems, there has been extensive work on devel-
oping dependent rounding techniques. Roughly speaking, these techniques round the fractional
solution in some random but correlated way to satisfy the hard constraints and also ensure some
concentration properties for the soft constraints. Some examples of such methods include swap
rounding [15, 16], randomized pipage rounding [2, 33, 19, 20], maximum-entropy sampling
[3, 32, 4], rounding via discrepancy [26, 29, 11] and Gaussian random walks [28].
    A key idea here is that the weaker property of negative dependence (instead of independence)
also suffices to get concentration. There is a rich and deep theory of negative dependence
and various notions such as negative correlation, negative cylinder dependence, negative
association, strongly Rayleigh property and determinantal measures, that imply interesting
concentration properties [27, 13, 17]. This insight has been extremely useful and for many
general problems such as those involving assignment or matroid polytopes, one can exploit
the underlying combinatorial structure to design rounding approaches that ensure negative
dependence between all or some suitable collection of random variables.


Limitations. Even though these dependent rounding methods are powerful and ingenious,
they are also limited by the fact that requiring negative dependence substantially restricts the
kinds of rounding steps that can be designed, and the type of problems that they can be applied
to. Moreover, even when such a rounding is possible, it typically requires a lot of creativity and
careful understanding of the problem structure to come up with the rounding for the problem
at hand.


1.2   Our results

Our main result is a new and general dependent rounding approach that we call sub-isotropic
rounding. In particular, it combines iterated and randomized rounding in a completely generic
way and significantly extends the scope of previous dependent rounding techniques. Before
describing our result, we need some definitions.
    Let ๐‘‹1 , . . . , ๐‘‹๐‘› be independent 0-1 random variables with means ๐‘ฅ ๐‘– = ๐”ผ[๐‘‹๐‘– ] and ๐‘Ž 1 , . . . , ๐‘Ž ๐‘›
be arbitrary reals (possibly negative). It is well known [14] that the sum ๐‘† = ๐‘– ๐‘Ž ๐‘– ๐‘‹๐‘– satisfies the
                                                                              ร


                       T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                            3
                                                N IKHIL BANSAL

following tail bound for any ๐‘ก โ‰ฅ 0:
                                                                                                    !
                                                                                    ๐‘ก2
      (Bernsteinโ€™s inequality)          Pr[๐‘† โˆ’ ๐”ผ[๐‘†] โ‰ฅ ๐‘ก] โ‰ค exp โˆ’ ร 2                                        (1.1)
                                                                2 ๐‘– ๐‘Ž ๐‘– (๐‘ฅ ๐‘– โˆ’ ๐‘ฅ 2๐‘– ) + 2๐‘€๐‘ก/3

where ๐‘€ = max๐‘– |๐‘Ž ๐‘– |. The lower tail follows by applying the above to โˆ’๐‘Ž ๐‘– , and the standard
Chernoff bounds correspond to (1.1) when ๐‘Ž ๐‘– โˆˆ [0, 1] for ๐‘– โˆˆ [๐‘›].
   The following relaxation of Bernsteinโ€™s inequality will be highly relevant for us.

Definition 1.1 (๐›ฝ-concentration). Let ๐›ฝ โ‰ฅ 1. For a vector-valued random variable ๐‘‹ =
(๐‘‹1 , . . . , ๐‘‹๐‘› ) where ๐‘‹๐‘– are possibly dependent 0-1 random variables, we say that ๐‘‹ is ๐›ฝ-
concentrated around its mean ๐‘ฅ = (๐‘ฅ 1 , . . . , ๐‘ฅ ๐‘› ) where ๐‘ฅ ๐‘– = ๐”ผ[๐‘‹๐‘– ], if for every ๐‘Ž โˆˆ โ„ ๐‘› , the
real-valued random variable h๐‘Ž, ๐‘‹i satisfies Bernsteinโ€™s inequality up to a factor of ๐›ฝ in the
exponent, i. e.,
                                                                                            !
                                                                          ๐‘ก 2 /๐›ฝ
                    Pr[h๐‘Ž, ๐‘‹i โˆ’ ๐”ผ[h๐‘Ž, ๐‘ฅi] โ‰ฅ ๐‘ก] โ‰ค exp โˆ’ ร 2                                                  (1.2)
                                                      2 ๐‘– ๐‘Ž ๐‘– (๐‘ฅ ๐‘– โˆ’ ๐‘ฅ 2๐‘– ) + 2๐‘€๐‘ก/3

where ๐‘€ = max๐‘– |๐‘Ž ๐‘– |.

Main result. We show that whenever iterated rounding can be applied to a problem such
that in iteration ๐‘˜, there is some slack in the sense that rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ for some fixed
๐›ฟ > 0, then ๐‘‚(1/๐›ฟ)-concentration can be achieved for free. More precisely, we have the following
result.

Theorem 1.2. For any fixed ๐›ฟ โˆˆ (0, 1), let us formalize an iterated rounding algorithm as follows. Given
a starting solution ๐‘ฅ, initialize ๐‘ฅ (0) = ๐‘ฅ. In each step ๐‘˜, for ๐‘˜ โ‰ฅ 1, the algorithm selects a matrix ๐‘Š (๐‘˜) with
rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ . Now it can pick any ๐‘ฆ (๐‘˜) satisfying ๐‘Š (๐‘˜) ๐‘ฆ (๐‘˜) = 0 and set ๐‘ฅ (๐‘˜) = ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜) ,
and iterate until ๐‘ฅ (final) is in {0, 1} ๐‘› . Let ๐‘‰ โŠ‚ {0, 1} ๐‘› be the set of outcomes that can be reached by the
iterated rounding algorithm.
     Then the sub-isotropic rounding algorithm outputs a random vector ๐‘‹ satisfying

   1. ๐‘‹ โˆˆ ๐‘‰ with probability 1, and

   2. ๐”ผ[๐‘‹] = ๐‘ฅ and ๐‘‹ is ๐›ฝ-concentrated around ๐‘ฅ with ๐›ฝ = 20/๐›ฟ.

Remark 1.3. A simple example shows that the dependence ๐›ฝ = ฮฉ(1/๐›ฟ) in Theorem 1.2 cannot
be improved. Let ๐›ฟ = 1/๐‘ก for some integer ๐‘ก, and consider ๐‘› variables ๐‘ฅ 1 , . . . , ๐‘ฅ ๐‘› , partitioned
into ๐‘›/๐‘ก blocks ๐ต1 , . . . , ๐ต๐‘›/๐‘ก where block ๐ต ๐‘– = {๐‘ฅ (๐‘–โˆ’1)๐‘ก+1 , . . . , ๐‘ฅ ๐‘–๐‘ก }. For each ๐ต ๐‘– there are ๐‘ก โˆ’ 1
constraints ๐‘ฅ(๐‘–โˆ’1)๐‘ก+1 = ๐‘ฅ (๐‘–โˆ’1)๐‘ก+2 = ยท ยท ยท = ๐‘ฅ ๐‘–๐‘ก , and hence there are (๐‘ก โˆ’ 1)(๐‘›/๐‘ก) = ๐‘›(1 โˆ’ ๐›ฟ) constraints
in total. Consider the starting solution with all ๐‘ฅ ๐‘— = 1/2. Now, no matter what random choices
the algorithm makes, the variables within a block evolve identically and all reach the same value
0 or 1. So the linear function ๐‘† = ๐‘ฅ 1 + ยท ยท ยท + ๐‘ฅ ๐‘› will only be 1/๐›ฟ-concentrated.

                         T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                   4
                 O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

    The generality of Theorem 1.2 directly gives new results for several problems where iterated
rounding gives useful guarantees. All one needs to show is that the original iterated rounding
argument for the problem can be applied with some slack, which is often straightforward and
only worsens the approximation guarantee slightly. In particular, note that Theorem 1.2 makes
no assumption about the combinatorial structure of the problem and by working with the more
relaxed notion of ๐›ฝ-concentration, we can completely avoid the need for negative dependence.

1.3     Motivating problems and applications
We now describe several applications of our result and also briefly discuss why they seem
beyond the reach of current dependent rounding methods.

1.3.1    Rounding for column-sparse LPs
Let ๐‘ฅ โˆˆ [0, 1]๐‘› be some fractional solution satisfying ๐ด๐‘ฅ = ๐‘, where ๐ด โˆˆ โ„ ๐‘šร—๐‘› is an ๐‘š ร— ๐‘›
matrix. The celebrated Beckโ€“Fiala algorithm [12] (see also [21] for a related result) uses iterated
rounding to produce an integral solution ๐‘‹ satisfying k๐ด(๐‘‹ โˆ’ ๐‘ฅ)k โˆž < ๐‘ก, where ๐‘ก is the maximum
โ„“ 1 norm of the columns of ๐ด. This is substantially better than randomized rounding for small ๐‘ก,
where the error for any row is typically its โ„“ 2 norm which can be substantially larger than ๐‘ก.
     Many problems, however, involve both some column-sparse constraints that come from the
underlying combinatorial problem, and some general arbitrary constraints which might not
have much structure. This motivates the following natural question.
Question 1.4. Let ๐‘€ be a linear system with two sets of constraints given by matrices ๐ด and ๐ต,
where ๐ด is column-sparse, while ๐ต is arbitrary. Given some fractional solution ๐‘ฅ, can we round
it to get error ๐‘‚(๐‘ก) for the rows of ๐ด, while doing no worse than randomized rounding for the
constraints in ๐ต?
Remark 1.5. Note that simply applying iterated rounding on the rows of ๐ด gives no control on
the error for ๐ต. Similarly, just doing randomized rounding will not give ๐‘‚(๐‘ก) error for ๐ด. Also
as ๐ด and ๐ต are arbitrary, negative dependence based techniques do not seem to apply.
     We show that a direct modification of the Beckโ€“Fiala argument gives slack ๐›ฟ, for any
๐›ฟ โˆˆ [0, 1), while worsening the error bound slightly to ๐‘ก/(1 โˆ’ ๐›ฟ). Setting, say ๐›ฟ = 1/2 and
applying Theorem 1.2 gives ๐‘‹ โˆˆ {โˆ’1, 1} ๐‘› that (i) has error at most 2๐‘ก for rows of ๐ด, (ii) satisfies
๐”ผ[๐‘‹๐‘– ] = ๐‘ฅ ๐‘– and is ๐‘‚(1)-concentrated, thus giving similar guarantees as randomized rounding
for the rows of ๐ต. In fact, the solution produced by the algorithm will satisfy concentration for
all linear constraints and not just for the rows of ๐ต.
     We also consider an extension to the Komlรณs setting, where the error depends on the
maximum โ„“2 norm of columns of ๐ด. These results are described in Section 4.1.

1.3.2    Makespan minimization on unrelated machines
The classical problem of makespan minimization on unrelated machines is the following. Given
๐‘› jobs and ๐‘š machines, where each job ๐‘— โˆˆ [๐‘›] has arbitrary size ๐‘ ๐‘–๐‘— on machine ๐‘– โˆˆ [๐‘š], assign

                      T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                        5
                                              N IKHIL BANSAL

the jobs to machines to minimize the maximum machine load. In a celebrated result, [24] gave
a rounding method with additive error at most ๐‘ max := max๐‘–๐‘— ๐‘ ๐‘–๐‘— , i. e., it gives an assignment
with makespan at most Opt + ๐‘ max where Opt is the value of an optimum LP solution. In
many practical problems, however, there are other soft resource constraints and side constraints
that are added to the fractional formulation. So it is useful to find a rounding that satisfies
these approximately but increases the makespan by only ๐‘‚(๐‘ max ). This motivates the following
natural problem.

Question 1.6. Given a fractional assignment ๐‘ฅ, find an integral assignment ๐‘‹ with additive
error ๐‘‚(๐‘max ) and that also satisfies ๐”ผ[๐‘‹๐‘–๐‘— ] = ๐‘ฅ ๐‘–๐‘— and concentration for all linear functions of ๐‘ฅ ๐‘–๐‘— ,
i. e., for all {๐‘Ž ๐‘–๐‘— } ๐‘–๐‘— , with high probability it holds that ๐‘–๐‘— ๐‘Ž ๐‘–๐‘— ๐‘‹๐‘–๐‘— โ‰ˆ ๐‘–๐‘— ๐‘Ž ๐‘–๐‘— ๐‘ฅ ๐‘–๐‘— .
                                                               ร             ร

    Questions related to finding a good assignment with some concentration properties have
been studied before [19, 2, 16], and several methods such as randomized pipage rounding and
swap rounding have been developed for this. However, these methods crucially rely on the
underlying matching structure and round the variables alternately along cycles, which limits
them in various ways: either they give partial assignments, or only get concentration for edges
incident to a vertex.
    We show that the iterated rounding proof of the result of [24] can be easily modified
to work for any slack ๐›ฟ โˆˆ [0, 1/2) while giving additive error ๐‘ max /(1 โˆ’ 2๐›ฟ). Theorem 1.2
(say, with ๐›ฟ = 1/4), thus gives a solution that has additive error at most 2๐‘ max and satisfies
๐‘‚(1)-concentration. The result also extends naturally to the ๐‘˜-resource setting, where ๐‘ ๐‘–๐‘— is a
๐‘˜-dimensional vector. These results are described in Section 4.2.

1.3.3   Degree-bounded spanning trees and thin trees
In the minimum cost degree-bounded spanning tree problem, we are given an undirected graph
๐บ = (๐‘‰ , ๐ธ) with edge costs ๐‘ ๐‘’ โ‰ฅ 0 for ๐‘’ โˆˆ ๐ธ, and integer degree bounds ๐‘ ๐‘ฃ for ๐‘ฃ โˆˆ ๐‘‰, and the goal
is to find a minimum cost spanning tree satisfying the degree bounds. In a breakthrough result,
Singh and Lau [31] gave an iterated rounding algorithm that given any fractional spanning tree
๐‘ฅ, finds a spanning tree ๐‘‡ with cost at most h๐‘, ๐‘ฅi and a degree violation of plus one.
     The celebrated thin-tree conjecture asks1 if given a fractional spanning tree ๐‘ฅ, there is a
spanning tree ๐‘‡ satisfying ฮ”๐‘‡ (๐‘†) โ‰ค ๐›ฝฮ”๐‘ฅ (๐‘†) for every ๐‘† โŠ‚ ๐‘‰, where ๐›ฝ = ๐‘‚(1). Here ฮ”๐‘‡ (๐‘†) is
the number of edges of ๐‘‡ crossing ๐‘†, and ฮ”๐‘ฅ (๐‘†) is the ๐‘ฅ-value crossing ๐‘†. This conjecture has
received a lot of attention recently, due to its connection to the asymmetric travelling salesman
problem (ATSP) [3, 1]. Despite the recent breakthrough on ATSP [34], the thin-tree conjecture
remains open.
     If one only considers single vertex sets ๐‘† = {๐‘ฃ}, the result of [31] implies that ฮ”๐‘‡ (๐‘ฃ) โ‰ค 2ฮ”๐‘ฅ (๐‘ฃ)
for each vertex ๐‘ฃ (as ฮ”๐‘ฅ (๐‘ฃ) โ‰ฅ 1 in any fractional spanning tree ๐‘ฅ). On the other hand for general
sets ๐‘†, the best known algorithmic results give ๐›ฝ = ๐‘‚(log ๐‘›/log log ๐‘›) [3, 15, 32, 20]. These
algorithms crucially rely on the negative dependence properties of spanning trees, which do
   1Equivalently, any ๐‘˜-edge-connected graph ๐บ has a spanning tree satisfying ฮ”๐‘‡ (๐‘†) = ๐‘‚(1/๐‘˜)ฮ”๐บ (๐‘†) for every
๐‘† โŠ‚ ๐‘‰.


                        T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                               6
                   O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

not give anything better for single vertex cuts (e. g., even if ๐‘ ๐‘ฃ = 2 for all ๐‘ฃ, by a balls-and-bins
argument a random tree will have maximum degree ฮ˜(log ๐‘›/log log ๐‘›)).
   The motivates the following natural question as a first step toward the thin-tree conjecture.

Question 1.7. Can we find a spanning tree that achieves ๐›ฝ = ๐‘‚(1) for single vertex cuts and
๐›ฝ = ๐‘‚(log ๐‘›/log log ๐‘›) for general cuts?

    We show that the iterated rounding algorithm of [31] can be easily modified to create slack
๐›ฟ โˆˆ (0, 1/2) while violating the degree bounds additively by less than 2/(1 โˆ’ 2๐›ฟ). Applying
Theorem 1.2 with ๐›ฟ = 1/6, the degree violation is strictly below 3 and this gives a distribution
supported on trees with a degree violation of plus 2 and satisfies ๐‘‚(1)-concentration. By a
standard cut counting argument [3], the concentration property implies ๐‘‚(log ๐‘›/log log ๐‘›)-
thinness for every cut. We describe these results in Section 4.3. In fact, we consider the more
general setting of the minimum cost degree-bounded matroid-basis problem.

1.4   Overview of techniques
We now give a high level overview of our algorithm and analysis. The starting observation is
that randomized rounding can be viewed as an iterative algorithm by doing a discrete version
of the standard Brownian motion on the cube as follows. Given ๐‘ฅ (0) as the starting fractional
solution, consider a random walk in the [0, 1]๐‘› cube starting at ๐‘ฅ (0) , with tiny step size ยฑ๐›พ
chosen independently for each coordinate, where upon reaching a face of the cube (i. e., some
๐‘ฅ ๐‘– reaches 0 or 1) the walk stays on that face. The process stops upon reaching some vertex
๐‘‹ = (๐‘‹1 , . . . , ๐‘‹๐‘› ) of the cube. By the martingale property of random walks, the probability
                            (0)
that ๐‘‹๐‘– = 1 is exactly ๐‘ฅ ๐‘– and as the walk in each coordinate is independent, ๐‘‹ has the same
distribution on {0, 1} ๐‘› as under randomized rounding.

A first attempt. Now consider iterated rounding, and recall that here the update ๐‘ฆ (๐‘˜) at iteration
๐‘˜ must lie in the nullspace of ๐‘Š (๐‘˜) . So a natural first idea to combine this with randomized
rounding, is to do a random walk in the null space of ๐‘Š (๐‘˜) until some variable reaches 0 or 1. The
slack condition rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ implies that the nullspace has at least ๐›ฟ๐‘› ๐‘˜ dimensions,
which could potentially give โ€œenough randomness" to the random walk.
    It turns out, however, that doing a standard random walk in the null space of ๐‘Š (๐‘˜) does
not work. The problem is that as the constraints defining ๐‘Š (๐‘˜) can be completely arbitrary in
our setting, the random walk can lead to very high correlations between certain subsets of
coordinates causing the ๐›ฝ-concentration property to fail. For example, suppose ๐›ฟ = 1/2 and ๐‘Š (0)
consists of the constraints ๐‘ฅ ๐‘– = ๐‘ฅ ๐‘–+1 for ๐‘– = 1, . . . , ๐‘›/2 โˆ’ 1. Then the random walk will update
๐‘ฅ ๐‘›/2 , . . . , ๐‘ฅ ๐‘› independently, but for ๐‘ฅ 1 , . . . , ๐‘ฅ ๐‘›/2 the updates must satisfy ฮ”๐‘ฅ1 = . . . = ฮ”๐‘ฅ ๐‘›/2 , and
hence will be completely correlated. So the linear function ๐‘ฅ 1 + . . . + ๐‘ฅ ๐‘›/2 will not have any
concentration (as all the variables will simultaneously rise by โˆ’๐›ฟ or by +๐›ฟ).

A different random walk. To get around this problem, we design a different random walk
in the null space of ๐‘Š (๐‘˜) , which looks similar to an independent walk in every direction even

                         T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                   7
                                                  N IKHIL BANSAL

though the coordinates are correlated. More formally, consider a random vector ๐‘Œ = (๐‘Œ1 , . . . , ๐‘Œ๐‘› ),
where ๐‘Œ๐‘– are mean-zero random variables. For a parameter ๐œ‚ โ‰ฅ 1, we say that ๐‘Œ is ๐œ‚-weakly
pairwise independent if for every ๐‘Ž = (๐‘Ž 1 , . . . , ๐‘Ž ๐‘› ) โˆˆ โ„ ๐‘› ,
                                                hร•          i    ร•
                                ๐”ผ[h๐‘Ž, ๐‘Œi 2 ] = ๐”ผ ( ๐‘Ž ๐‘– ๐‘Œ๐‘– )2 โ‰ค ๐œ‚   ๐‘Ž 2๐‘– ๐”ผ[๐‘Œ๐‘–2 ].
                                                      ๐‘–                  ๐‘–

If ๐‘Œ1 , . . . , ๐‘Œ๐‘› are pairwise independent, note that the above holds as equality with ๐œ‚ = 1, and
hence this can be viewed as a relaxation of pairwise independence. We show that whenever
rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ , there exist ๐œ‚-weakly pairwise independent random updates ๐‘ฆ (๐‘˜) that
lie in the null space of ๐‘Š (๐‘˜) (which has dimension at least ๐›ฟ๐‘› ๐‘˜ ) with ๐œ‚ โ‰ˆ 1/๐›ฟ. Moreover these
updates can be found by solving a semidefinite program (SDP).
     Next, using a variant of Freedmanโ€™s martingale analysis [18], we show that applying these
๐œ‚-weakly pairwise independent random updates (with small increments) until all the variables
reach 0-1, gives an integral solution that satisfies ๐‘‚(๐œ‚)-concentration.
     These techniques are motivated by our recent works on algorithmic discrepancy [8, 9]. While
discrepancy is closely related to rounding [25, 29], a key difference in discrepancy is that the
error for rounding a linear system ๐ด๐‘ฅ = ๐‘ depends on the โ„“ 2 norms of the coefficients of the
constraints and not on ๐‘. E. g., suppose ๐‘ฅ โˆˆ [0, 1]๐‘› satisfies ๐‘ฅ 1 + . . . + ๐‘ฅ ๐‘› = log ๐‘›, then the
sum stays ๐‘‚(log ๐‘›) upon randomized
                                โˆš         rounding with high probability, while using discrepancy
methods directly gives ฮฉ( ๐‘›) error, which would be unacceptably large in this setting. So our
results can be viewed as using techniques from discrepancy theory to obtain bounds that are
sensitive to ๐‘ฅ. Recently, this direction was explored in [11] but their method gave much weaker
results and applied to very limited settings.


2     Technical preliminaries
2.1     Tail bounds for supermartingales
We will need the following tail bound for supermartingales with a strong negative drift.
Theorem 2.1. Let ๐›ผ โˆˆ (0, 1). Let {๐‘ ๐‘˜ : ๐‘˜ = 0, 1, . . . , } be a sequence of random variables with
๐‘Œ๐‘˜ := ๐‘ ๐‘˜ โˆ’ ๐‘ ๐‘˜โˆ’1 , such that ๐‘0 is constant and ๐‘Œ๐‘˜ โ‰ค 1 for all ๐‘˜ โ‰ฅ 1. If

                                             ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜ ] โ‰ค โˆ’๐›ผ ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜2 ]

for all ๐‘˜ โ‰ฅ 1, where ๐”ผ ๐‘˜โˆ’1 [ ยท ] denotes ๐”ผ[ ยท |๐‘1 , . . . , ๐‘ ๐‘˜โˆ’1 ], then for ๐‘ก โ‰ฅ 0,

                                          Pr[๐‘ ๐‘˜ โˆ’ ๐‘0 โ‰ฅ ๐‘ก] โ‰ค exp(โˆ’๐›ผ๐‘ก).

      Before proving Theorem 2.1, we first need a simple lemma.
Lemma 2.2. Let ๐‘‹ be a random variable satisfying ๐‘‹ โ‰ค 1. Then for any ๐œ† > 0,
                                                                                      
                                    ๐œ†๐‘‹                          ๐œ†
                                ๐”ผ[e      ] โ‰ค exp ๐œ†๐”ผ[๐‘‹] + (e โˆ’ ๐œ† โˆ’ 1)๐”ผ[๐‘‹ ] .       2



                          T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                      8
                    O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

Proof. Let ๐‘“ (๐œ†, ๐‘ฅ) = (e๐œ†๐‘ฅ โˆ’ ๐œ†๐‘ฅ โˆ’ 1)/๐‘ฅ 2 , where we set ๐‘“ (๐œ†, 0) = ๐œ†2 /2. By standard integration, it
                                    โˆซ๐œ†โˆซ๐‘ 
is easily verified that ๐‘“ (๐œ†, ๐‘ฅ) = 0 0 e๐‘ก๐‘ฅ ๐‘‘๐‘ก ๐‘‘๐‘ . As e๐‘ก๐‘ฅ is non-decreasing in ๐‘ฅ for all ๐‘ก โ‰ฅ 0, this
implies that ๐‘“ (๐œ†, ๐‘ฅ) is non-decreasing in ๐‘ฅ. In particular, ๐‘“ (๐œ†, ๐‘ฅ) โ‰ค ๐‘“ (๐œ†, 1) for any ๐‘ฅ โ‰ค 1 and
hence
             e๐œ†๐‘ฅ = 1 + ๐œ†๐‘ฅ + ๐‘“ (๐œ†, ๐‘ฅ)๐‘ฅ 2 โ‰ค 1 + ๐œ†๐‘ฅ + ๐‘“ (๐œ†, 1)๐‘ฅ 2 = 1 + ๐œ†๐‘ฅ + (e๐œ† โˆ’ ๐œ† โˆ’ 1)๐‘ฅ 2 .
Taking expectations and using that 1 + ๐‘ฅ โ‰ค e๐‘ฅ for all ๐‘ฅ โˆˆ โ„ gives,
                                                                                             
              ๐œ†๐‘‹                     ๐œ†                                      ๐œ†
          ๐”ผ[e      ] โ‰ค 1 + ๐œ†๐”ผ[๐‘‹] + (e โˆ’ ๐œ† โˆ’ 1)๐”ผ[๐‘‹ ] โ‰ค exp ๐œ†๐”ผ[๐‘‹] + (e โˆ’ ๐œ† โˆ’ 1)๐”ผ[๐‘‹ ] .
                                                     2                                    2
                                                                                                      

Proof of Theorem 2.1. By Markovโ€™s inequality,

                                                                       ๐”ผ[exp(๐›ผ(๐‘ ๐‘˜ โˆ’ ๐‘0 ))]
             Pr[๐‘ ๐‘˜ โˆ’ ๐‘0 โ‰ฅ ๐‘ก] = Pr[exp(๐›ผ(๐‘ ๐‘˜ โˆ’ ๐‘0 )) โ‰ฅ exp(๐›ผ๐‘ก)] โ‰ค
                                                                                exp(๐›ผ๐‘ก)

so it suffices to show that ๐”ผ[exp(๐›ผ(๐‘ ๐‘˜ โˆ’ ๐‘0 ))] โ‰ค 1. As ๐‘0 is deterministic, this is same as
๐”ผ[exp(๐›ผ๐‘ ๐‘˜ )] โ‰ค exp(๐›ผ๐‘0 ). Now,
                                              h              i
                  ๐”ผ ๐‘˜โˆ’1 e๐›ผ๐‘ ๐‘˜ = e๐›ผ๐‘ ๐‘˜โˆ’1 ๐”ผ ๐‘˜โˆ’1 e๐›ผ(๐‘ ๐‘˜ โˆ’๐‘ ๐‘˜โˆ’1 ) = e๐›ผ๐‘ ๐‘˜โˆ’1 ๐”ผ ๐‘˜โˆ’1 e๐›ผ๐‘Œ๐‘˜
                                                                                
                                                                 
                โ‰ค e๐›ผ๐‘ ๐‘˜โˆ’1 exp ๐›ผ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜ ] + (e๐›ผ โˆ’ ๐›ผ โˆ’ 1)๐”ผ ๐‘˜โˆ’1 ๐‘Œ๐‘˜2             (Lemma 2.2)
                                                              
                โ‰ค e๐›ผ๐‘ ๐‘˜โˆ’1 exp (e๐›ผ โˆ’ ๐›ผ2 โˆ’ ๐›ผ โˆ’ 1)๐”ผ ๐‘˜โˆ’1 ๐‘Œ๐‘˜2
                                                         

                โ‰ค e๐›ผ๐‘ ๐‘˜โˆ’1      (as e๐›ผ โ‰ค 1 + ๐›ผ + ๐›ผ2 for 0 โ‰ค ๐›ผ โ‰ค 1).

As this holds for all ๐‘˜, the result follows by the property of Iterated Expectations.                 

2.2   Semidefinite matrices
Let ๐‘€๐‘› denote the class of all symmetric ๐‘› ร— ๐‘› matrices with real entries. For two matrices
๐ด, ๐ต โˆˆ โ„ ๐‘›ร—๐‘› , the trace inner product of ๐ด and ๐ต is defined as h๐ด, ๐ตi = tr(๐ด๐‘‡ ๐ต) = ๐‘›๐‘–=1 ๐‘›๐‘—=1 ๐‘Ž ๐‘–๐‘— ๐‘ ๐‘–๐‘— .
                                                                                     ร ร
A matrix ๐‘ˆ โˆˆ ๐‘€๐‘› is positive semidefinite (PSD) if all its eigenvalues are non-negative and we
denote this by ๐‘ˆ  0. Equivalently, ๐‘ˆ  0 iff ๐‘Ž ๐‘‡ ๐‘ˆ ๐‘Ž = h๐‘Ž๐‘Ž ๐‘‡ , ๐‘ˆi โ‰ฅ 0 for all ๐‘Ž โˆˆ โ„ ๐‘› .
    For, ๐‘ˆ  0 let ๐‘ˆ 1/2 = ๐‘– ๐œ† ๐‘– ๐‘ข๐‘– ๐‘ข๐‘–๐‘‡ , where ๐‘ˆ = ๐‘– ๐œ† ๐‘– ๐‘ข๐‘– ๐‘ข๐‘–๐‘‡ is the spectral decomposition of ๐‘ˆ
                            ร 1/2                    ร

with orthonormal eigenvectors ๐‘ข๐‘– . Then ๐‘ˆ 1/2 is PSD and ๐‘ˆ = ๐‘‰ ๐‘‡ ๐‘‰ for ๐‘‰ = ๐‘ˆ 1/2 . For ๐‘Œ, ๐‘ โˆˆ ๐‘€๐‘› ,
we say that ๐‘Œ  ๐‘ if ๐‘ โˆ’ ๐‘Œ  0.

2.3   Approximate independence and sub-isotropic random variables
Let ๐‘Œ = (๐‘Œ1 , . . . , ๐‘Œ๐‘› ) be a random vector with ๐‘Œ1 , . . . , ๐‘Œ๐‘› possibly dependent.

Definition 2.3 ((๐œ, ๐œ‚) sub-isotropic random vector). For ๐œ โˆˆ (0, 1] and ๐œ‚ โ‰ฅ 1, we say that ๐‘Œ is
(๐œ, ๐œ‚) sub-isotropic if it satisfies the following conditions.

                        T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                           9
                                                  N IKHIL BANSAL

                                                                      ร๐‘›
   1. ๐”ผ[๐‘Œ๐‘– ] = 0 and ๐”ผ[๐‘Œ๐‘–2 ] โ‰ค 1 for all ๐‘– โˆˆ [๐‘›], and                  ๐‘–=1 ๐”ผ[๐‘Œ๐‘– ] โ‰ฅ ๐œ๐‘›.
                                                                               2



   2. For all ๐‘ = (๐‘ 1 , . . . , ๐‘ ๐‘› ) โˆˆ โ„ ๐‘› it holds that

                                              ๐‘›
                                             hร•                   i        ๐‘›
                                                                           ร•
                                           ๐”ผ (         ๐‘ ๐‘– ๐‘Œ๐‘– )
                                                              2
                                                                      โ‰ค๐œ‚         ๐‘ 2๐‘– ๐”ผ[๐‘Œ๐‘– ]2 .      (2.1)
                                                 ๐‘–=1                       ๐‘–=1



   Note that if ๐‘Œ1 , . . . , ๐‘Œ๐‘› are pairwise independent then (2.1) holds with equality for ๐œ‚ = 1.
   Let ๐‘ˆ โˆˆ ๐‘€๐‘› be the ๐‘› ร— ๐‘› covariance matrix of ๐‘Œ1 , . . . , ๐‘Œ๐‘› , i. e., ๐‘ˆ ๐‘–๐‘— = ๐”ผ[๐‘Œ๐‘– ๐‘Œ๐‘— ]. Every covariance
matrix is PSD as ๐‘ ๐‘‡ ๐‘ˆ ๐‘ = ๐”ผ[( ๐‘– ๐‘ ๐‘– ๐‘Œ๐‘– )2 ] โ‰ฅ 0 for all ๐‘ โˆˆ โ„ ๐‘› . Let diag(๐‘ˆ) denote the diagonal ๐‘› ร— ๐‘›
                                  ร
matrix with entries ๐‘ˆ ๐‘–๐‘– , then (2.1) can be written as ๐‘ ๐‘‡ (๐œ‚ diag(๐‘ˆ) โˆ’ ๐‘ˆ)๐‘ โ‰ฅ 0 for every ๐‘ โˆˆ โ„ ๐‘› ,
and hence equivalently expressed as

                                                  ๐‘ˆ  ๐œ‚ diag(๐‘ˆ).


Generic construction. We will use the following generic way to produce (๐œ, ๐œ‚) sub-isotropic
vectors. Let ๐‘ˆ be a ๐‘› ร— ๐‘› PSD matrix satisfying: ๐‘ˆ ๐‘–๐‘– โ‰ค 1, tr(๐‘ˆ) โ‰ฅ ๐œ๐‘› and ๐‘ˆ  ๐œ‚ diag(๐‘ˆ). Let
๐‘Ÿ โˆˆ โ„ ๐‘› be a random vector where each coordinate is independently and uniformly ยฑ1. Then the
random vector ๐‘Œ = ๐‘ˆ 1/2 ๐‘Ÿ has covariance ๐”ผ[๐‘Œ๐‘Œ ๐‘‡ ] = ๐‘ˆ 1/2 ๐”ผ[๐‘Ÿ๐‘Ÿ ๐‘‡ ](๐‘ˆ ๐‘‡ )1/2 = ๐‘ˆ, and the properties
of ๐‘ˆ imply that ๐‘Œ is (๐œ, ๐œ‚) sub-isotropic.

Remark 2.4. In other similar constructions, ๐‘Ÿ is often chosen as a random Gaussian, but we
prefer to choose it as random ยฑ1 as it is bounded, and this makes some technical arguments
easier later on.

   We will need the following result from [9], about finding sub-isotropic random vectors
orthogonal to a subspace.

Theorem 2.5 ([9], Theorem 6). Let ๐บ โŠ‚ โ„ ๐‘› be an arbitrary subspace with dimension dim(๐บ) = โ„“ = ๐›ฟ๐‘›.
Then for any ๐œ > 0 and ๐œ‚ > 1 satisfying 1/๐œ‚ + ๐œ โ‰ค ๐›ฟ, there is a ๐‘› ร— ๐‘› PSD matrix ๐‘ˆ, that is computable
in polynomial time, and satisfies the following properties:
    (i) hโ„Ž โ„Ž ๐‘‡ , ๐‘ˆi = 0 for all โ„Ž โˆˆ ๐บ โŠฅ , where ๐บโŠฅ is the subspace orthogonal to ๐บ.
    (ii) ๐‘ˆ ๐‘–๐‘– โ‰ค 1 for all ๐‘– โˆˆ [๐‘›].
    (iii) tr(๐‘ˆ) โ‰ฅ ๐œ๐‘›.
    (iv) ๐‘ˆ  ๐œ‚ diag(๐‘ˆ).

    The first condition gives that the range space of ๐‘ˆ is contained in the subspace ๐บ, as
hโ„Ž โ„Ž ๐‘‡ , ๐‘ˆi = 0 implies that k๐‘ˆ 1/2 โ„Ž k 2 = 0 and hence โ„Ž ๐‘‡ ๐‘ˆ 1/2 = 0. So, for any vector ๐‘Ÿ โˆˆ โ„ ๐‘› the
vector ๐‘Œ = ๐‘ˆ 1/2 ๐‘Ÿ lies in ๐บ (as for every โ„Ž โˆˆ ๐บ โŠฅ , โ„Ž ๐‘‡ ๐‘Œ = โ„Ž ๐‘‡ ๐‘ˆ 1/2 ๐‘Ÿ = h0, ๐‘Ÿi = 0). So this gives a
polynomial time algorithm to generate a (๐œ, ๐œ‚) sub-isotropic random vector ๐‘Œ โˆˆ ๐บ.

                         T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                           10
                  O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

3     The algorithm and analysis
Recall that by iterated rounding we refer to any procedure that given some starting fractional
solution ๐‘ฅ, sets ๐‘ฅ (0) = ๐‘ฅ, and applies a sequence of updates as follows. Given the vector ๐‘ฅ (๐‘˜โˆ’1) at
                                                                    (๐‘˜โˆ’1)
the beginning of iteration ๐‘˜, call a variable ๐‘– โˆˆ [๐‘›] frozen if ๐‘ฅ ๐‘–       is 0 or 1, and alive otherwise.
Let ๐‘› ๐‘˜ denote the number of alive variables. Based on ๐‘ฅ          (๐‘˜โˆ’1) , the algorithm picks a set of
constraints of rank at most ๐‘› ๐‘˜ โˆ’ 1, given by the rows of some matrix ๐‘Š (๐‘˜) , and finds some
                                                       (๐‘˜)
non-zero vector ๐‘ฆ (๐‘˜) satisfying ๐‘Š (๐‘˜) ๐‘ฆ (๐‘˜) = 0 and ๐‘ฆ ๐‘– = 0 for ๐‘– frozen. The solution is updated as
๐‘ฅ (๐‘˜) = ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜) .
     Let ๐›ฟ > 0 be the slack parameter. We assume that the problem to be solved has an
iterated rounding procedure where in each iteration ๐‘˜ one can choose some matrix ๐‘Š (๐‘˜) with
rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ . We now describe the rounding algorithm.


3.1   The algorithm
Initialize ๐‘ฅ (0) = ๐‘ฅ, where ๐‘ฅ in the starting fractional solution given as input. For each iteration
๐‘˜ = 1, . . . , repeat the following until all the variables reach 0 or 1.

Iteration ๐‘˜. Let ๐‘ฅ (๐‘˜โˆ’1) be the solution at the beginning of iteration ๐‘˜, and let ๐ด ๐‘˜ โŠ‚ [๐‘›] denote the
                                  (๐‘˜โˆ’1)
set of alive coordinates ๐‘– with ๐‘ฅ ๐‘–     โˆˆ (0, 1). Only the variables in ๐ด ๐‘˜ will be updated henceforth.
So for ease of notation we assume that ๐ด ๐‘˜ = [๐‘› ๐‘˜ ].

    1. Apply Theorem 2.5, with ๐‘› = ๐‘› ๐‘˜ , ๐บ the nullspace of ๐‘Š (๐‘˜) , ๐œ = ๐›ฟ/10 and ๐œ‚ = 10/(9๐›ฟ) to find
       the covariance matrix ๐‘ˆ.

    2. Let ๐›พ = 1/(2๐‘› 3/2 ). Let ๐‘Ÿ ๐‘˜ โˆˆ โ„ ๐‘› ๐‘˜ be a random vector with independent ยฑ1 entries. Set

                               ๐‘ฅ (๐‘˜) = ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜)   with   ๐‘ฆ (๐‘˜) = ๐›พ๐‘˜ ๐‘ˆ 1/2 ๐‘Ÿ ๐‘˜ ,

      where ๐›พ๐‘˜ is the largest value such that (i) ๐›พ๐‘˜ โ‰ค ๐›พ and (ii) both ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜) and ๐‘ฅ (๐‘˜โˆ’1) โˆ’ ๐‘ฆ (๐‘˜)
      lie in [0, 1]๐‘› ๐‘˜ .


3.2   Analysis
Let ๐‘‹ = (๐‘‹1 , . . . , ๐‘‹๐‘› ) denote the final 0-1 solution returned by the algorithm. The property that
๐”ผ[๐‘‹๐‘– ] = ๐‘ฅ ๐‘– follows directly as the update ๐‘ฆ (๐‘˜) at each time step has mean zero in each coordinate.
As the algorithm always moves in the nullspace of ๐‘Š (๐‘˜) , clearly it will also satisfy the iterated
rounding guarantee with probability 1.
    To analyze the running time, we note that whenever ๐›พ๐‘˜ < ๐›พ, there is at least 1/2 probability
that some new variable will reach 0 or 1 after the update (as the new solution is either ๐‘ฅ (๐‘˜) + ๐‘ฆ (๐‘˜)
or ๐‘ฅ (๐‘˜) โˆ’ ๐‘ฆ (๐‘˜) with probability half each). So, in expectation there are at most 2๐‘› such steps. So
let us focus on the iterations where ๐›พ๐‘˜ = ๐›พ.

                      T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                             11
                                                      N IKHIL BANSAL

                                                                               (๐‘˜)
    Let us define the energy of ๐‘ฅ (๐‘˜) as ๐ธ (๐‘˜) := ๐‘– (๐‘ฅ ๐‘– )2 . During the update, conditioned on all
                                                                   ร

the randomness up to time ๐‘˜ โˆ’ 1, ๐ธ (๐‘˜) rises in expectation by at least ๐›พ 2 ๐‘› ๐‘˜ ๐œ as,
                                                 hร•                         i
                                                       (๐‘˜โˆ’1) (๐‘˜)     (๐‘˜)
                  ๐”ผ ๐‘˜โˆ’1 ๐ธ(๐‘˜) โˆ’ ๐ธ(๐‘˜โˆ’1) = ๐›พ2 ๐”ผ ๐‘˜โˆ’1            ๐‘ฆ ๐‘– + (๐‘ฆ ๐‘– )2 )
                                    
                                                   (2๐‘ฅ ๐‘–
                                                                           ๐‘–
                                                              ร•
                                                  = ๐›พ               ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘–(๐‘˜) )2 = ๐›พ2 tr(๐‘ˆ) โ‰ฅ ๐›พ2 ๐œ๐‘› ๐‘˜ .
                                                         2
                                                                                   
                                                               ๐‘–
                                                          (๐‘˜) 
where the second equality uses that ๐”ผ ๐‘˜โˆ’1 ๐‘ฆ ๐‘–   = 0 for each ๐‘–. As the final energy can be at most
๐‘›, standard arguments [6, 26] imply that, with constant probability, the algorithm terminates in
๐‘‚(๐‘› log ๐‘› + log ๐‘›/๐›พ2 ) time steps.
    It remains to show that the rounding satisfies the concentration property, which we do next.

3.2.1   Isotropic updates imply concentration
Let ๐‘‹ = (๐‘‹1 , . . . , ๐‘‹๐‘› ) denote the final 0-1 solution returned by the algorithm. Fix any ๐‘Ž =
(๐‘Ž 1 , ๐‘Ž2 , . . . , ๐‘Ž ๐‘› ) โˆˆ โ„ ๐‘› . We will show that
                                                                                                                           !
                                                                                                  ๐‘ก 2 /๐›ฝ
                   Pr[h๐‘Ž, ๐‘ฅi โˆ’ ๐”ผ[h๐‘Ž, ๐‘ฅi] โ‰ฅ ๐‘ก] โ‰ค exp โˆ’ ร 2
                                                     2( ๐‘– ๐‘Ž ๐‘– (๐‘ฅ ๐‘– โˆ’ ๐‘ฅ 2๐‘– ) + ๐‘€๐‘ก/3)
with ๐›ฝ = 18๐œ‚ which equals 20/๐›ฟ by our choice of the parameters, and ๐‘€ = max๐‘– |๐‘Ž ๐‘– |.

Proof. By scaling ๐‘Ž ๐‘– โ€™s and ๐‘ก, we can assume that ๐‘€ = 1. Let us define the random variable
                                             ร•                       ร•
                                                         (๐‘˜)                          (๐‘˜)         (๐‘˜)
                                   ๐‘๐‘˜ =             ๐‘Ž๐‘– ๐‘ฅ๐‘– + ๐œ†                   ๐‘Ž 2๐‘– ๐‘ฅ ๐‘– (1 โˆ’ ๐‘ฅ ๐‘– ),
                                              ๐‘–                        ๐‘–

where the parameter ๐œ† โ‰ค 1 will be optimized later.
                                                                                             (๐‘˜)      (๐‘˜)
     It is useful to think of ๐‘ ๐‘˜ as ๐œ‡ ๐‘˜ + ๐œ†๐‘ฃ ๐‘˜ , where ๐œ‡ ๐‘˜ = h๐‘Ž, ๐‘ฅ (๐‘˜) i and ๐‘ฃ ๐‘˜ = ๐‘– ๐‘Ž 2๐‘– ๐‘ฅ ๐‘– (1 โˆ’ ๐‘ฅ ๐‘– ) are
                                                                                   ร

the mean and variance if we would randomly round the coordinates of ๐‘ฅ (๐‘˜) to 0-1. Initially,
๐‘0 = ๐œ‡ + ๐œ†๐‘ฃ where ๐œ‡ = ๐‘– ๐‘Ž ๐‘– ๐‘ฅ ๐‘– and ๐‘ฃ = ๐‘– ๐‘Ž 2๐‘– (๐‘ฅ ๐‘– โˆ’ ๐‘ฅ 2๐‘– ).
                            ร                ร
     We will show that ๐‘ ๐‘˜ satisfies the conditions of Theorem 2.1 for an appropriate ๐›ผ, and use
it to show the required tail bound. Roughly speaking, as the algorithm proceeds, ๐‘ ๐‘˜ has a
strong negative drift as the energy term ๐‘ฃ ๐‘˜ decreases in expectation and ๐œ‡ ๐‘˜ does not change in
expectation. In the proof of Theorem 2.1, this negative drift offsets the positive terms that arise
while bounding the exponential moment of ๐‘ ๐‘˜ .
     We now give the details. We first compute ๐‘Œ๐‘˜ := ๐‘ ๐‘˜ โˆ’ ๐‘ ๐‘˜โˆ’1 and show that it is bounded.
Using ๐‘ฅ (๐‘˜) = ๐‘ฅ (๐‘˜โˆ’1) + ๐‘ฆ (๐‘˜) , we have
             ๐‘Œ๐‘˜ = ๐‘ ๐‘˜ โˆ’ ๐‘ ๐‘˜โˆ’1
                      ร•                                  ร•                                                                         
                                 (๐‘˜)        (๐‘˜โˆ’1)                               (๐‘˜)         (๐‘˜)         (๐‘˜โˆ’1)          (๐‘˜โˆ’1)
                  =         ๐‘Ž ๐‘– (๐‘ฅ ๐‘– โˆ’ ๐‘ฅ ๐‘–          )+             ๐œ†๐‘Ž 2๐‘– (๐‘ฅ ๐‘– (1 โˆ’ ๐‘ฅ ๐‘– ) โˆ’ ๐‘ฅ ๐‘–                  (1 โˆ’ ๐‘ฅ ๐‘–       ))
                        ๐‘–                                 ๐‘–
                      ร•                ร•                                                     
                                 (๐‘˜)                   (๐‘˜)       (๐‘˜โˆ’1)    (๐‘˜)
                  =         ๐‘Ž ๐‘– ๐‘ฆ๐‘– +           ๐œ†๐‘Ž 2๐‘– ๐‘ฆ ๐‘– (1 โˆ’ 2๐‘ฅ ๐‘–     โˆ’ ๐‘ฆ๐‘– )                     .                                     (3.1)
                        ๐‘–               ๐‘–


                       T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                                               12
                   O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

Claim 3.1. For all ๐‘˜, the update ๐‘ฆ (๐‘˜) satifies k๐‘ฆ (๐‘˜) k 2 โ‰ค ๐›พ ๐‘› = 2โˆš1 ๐‘› .

Proof. Recall that ๐‘ฆ (๐‘˜) = ๐›พ๐‘ˆ 1/2 ๐‘Ÿ ๐‘˜ . Let ๐‘ˆ 1/2 (๐‘–) denote the ๐‘–-th column of ๐‘ˆ 1/2 . As

                                             h๐‘ˆ 1/2 (๐‘–), ๐‘ˆ 1/2 (๐‘–)i = ๐‘ˆ ๐‘–๐‘– โ‰ค 1,

the columns of ๐‘ˆ 1/2 have length at most 1. Let ๐‘Ÿ ๐‘˜ (๐‘–) denote the ๐‘–-th entry of ๐‘Ÿ ๐‘˜ . Applying the
triangle inequality to the columns of ๐‘ˆ 1/2 ,
                                                    ร•
                                 k๐‘ˆ 1/2 ๐‘Ÿ ๐‘˜ k 2 โ‰ค            |๐‘Ÿ ๐‘˜ (๐‘–)|k๐‘ˆ 1/2 (๐‘–)k 2 โ‰ค k๐‘Ÿ ๐‘˜ k 1 โ‰ค ๐‘›.
                                                      ๐‘–


This gives that k๐‘ฆ (๐‘˜) k 2 โ‰ค ๐›พ๐‘›.                                                                                                             

Claim 3.2. For all ๐‘˜, |๐‘Œ๐‘˜ | โ‰ค 1.

                                                                            (๐‘˜)
                                                                   ๐‘– |๐‘Ž ๐‘– ๐‘ฆ ๐‘– |. This follows as |๐‘Ž ๐‘– | โ‰ค |๐‘Ž ๐‘– |
                                                                                             ร
Proof. First we note that the second term in (3.1) is at most                                          2
                                                     (๐‘˜โˆ’1)     (๐‘˜)                         (๐‘˜โˆ’1)
(as ๐‘€ = 1), ๐œ† โ‰ค 1 by our assumption, and 1 โˆ’ 2๐‘ฅ ๐‘–          โˆ’ ๐‘ฆ ๐‘– โˆˆ [โˆ’1, 1] (as 1 โˆ’ ๐‘ฅ ๐‘–           โˆˆ [0, 1] and
  (๐‘˜โˆ’1)     (๐‘˜)   (๐‘˜)
๐‘ฅ๐‘–      + ๐‘ฆ ๐‘– = ๐‘ฅ ๐‘– โˆˆ [0, 1]). As k๐‘Ž k โˆž โ‰ค 1 and using the bound on k๐‘ฆ k 2 in Claim 3.1, we have
                                                                                (๐‘˜)

                                            ร•                    ร•                           ร•
                                                       (๐‘˜)                       (๐‘˜)                   (๐‘˜)
                                  |๐‘Œ๐‘˜ | โ‰ค       ๐‘Ž ๐‘– ๐‘ฆ๐‘– +                   |๐‘Ž ๐‘– ๐‘ฆ ๐‘– | โ‰ค 2          |๐‘ฆ ๐‘– |
                                            ๐‘–                     ๐‘–                           ๐‘–
                                                (๐‘˜)                              (๐‘˜)
                                       = 2k๐‘ฆ          k 1 โ‰ค 2๐‘›        1/2
                                                                            k๐‘ฆ         k 2 โ‰ค 2๐›พ๐‘› 3/2 = 1.                                    

    We now upper bound the negative drift of ๐‘ ๐‘˜ .

Claim 3.3. ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜ ] โ‰ค โˆ’(๐œ†/8๐œ‚)๐”ผ ๐‘˜โˆ’1 ๐‘Œ๐‘˜2
                                                     

                  (๐‘˜)                                      (๐‘˜โˆ’1)
Proof. As ๐”ผ ๐‘˜โˆ’1 ๐‘ฆ ๐‘–   = 0 for all ๐‘–, and as ๐‘ฅ ๐‘–     is deterministic conditioned on the randomness
until ๐‘˜ โˆ’ 1, taking expectations ๐”ผ ๐‘˜โˆ’1 [ยท] in (3.1) gives
                                                                  ร•
                                                                                             (๐‘˜)
                                                                            ๐‘Ž 2๐‘– ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘– )2 .
                                                                                                  
                                        ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜ ] = โˆ’๐œ†                                                                                  (3.2)
                                                                       ๐‘–


We now upper bound ๐”ผ ๐‘˜โˆ’1 [๐‘Œ๐‘˜2 ]. Using (๐‘Ž + ๐‘)2 โ‰ค 2๐‘Ž 2 + 2๐‘ 2 twice for the expression in (3.1),
                                                                                                             !2
                            ร•                             ร•
                                        (๐‘˜)                             (๐‘˜)       (๐‘˜โˆ’1)    (๐‘˜)
             ๐‘Œ๐‘˜2 โ‰ค 2(             ๐‘Ž ๐‘– ๐‘ฆ ๐‘– )2 + 2๐œ†2               ๐‘Ž 2๐‘– ๐‘ฆ ๐‘– (1 โˆ’ 2๐‘ฅ ๐‘–     โˆ’ ๐‘ฆ๐‘– )
                             ๐‘–                               ๐‘–
                                                                                                   !2                            !2
                            ร•                      ยฉ ร• 2 (๐‘˜)                                                 ร•
                                        (๐‘˜)                          (๐‘˜โˆ’1)                                                 (๐‘˜)
                   โ‰ค 2(           ๐‘Ž ๐‘– ๐‘ฆ ๐‘– )2 + 4๐œ†2 ยญ   ๐‘Ž ๐‘– ๐‘ฆ (1 โˆ’ 2๐‘ฅ ๐‘–     )                            +             ๐‘Ž 2๐‘– (๐‘ฆ ๐‘– )2 ยฎ
                                                                                                                                      ยช
                                                                                                                                          (3.3)
                             ๐‘–                     ยซ ๐‘–                                                            ๐‘–                   ยฌ

                           T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                                             13
                                                               N IKHIL BANSAL

Taking expectations ๐”ผ ๐‘˜โˆ’1 [ยท] in (3.3), we now upper bound the terms on the right. As ๐‘ฆ (๐‘˜) is (๐œ, ๐œ‚)
sub-isotropic, by (2.1), the first term satisfies
                                         hร•           i    ร•
                                                  (๐‘˜)                   (๐‘˜) 
                                    ๐”ผ ๐‘˜โˆ’1 ( ๐‘Ž ๐‘– ๐‘ฆ ๐‘– )2 โ‰ค ๐œ‚   ๐‘Ž 2๐‘– ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘– )2 .
                                                   ๐‘–                        ๐‘–

Similarly, by the sub-isotropic property, the second term satisfies
           hร•                                  i           ร•                                                      ร•
                           (๐‘˜)       (๐‘˜โˆ’1) 2                                 (๐‘˜โˆ’1) 2                    (๐‘˜)                       (๐‘˜)
                    ๐‘Ž 2๐‘– ๐‘ฆ ๐‘– (1 โˆ’ 2๐‘ฅ ๐‘–             โ‰ค๐œ‚           ๐‘Ž 4๐‘– (1 โˆ’ 2๐‘ฅ ๐‘–                                        ๐‘Ž 2๐‘– ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘– )2 ,
                                                                                                                                     
      ๐”ผ ๐‘˜โˆ’1 (                             ))                                           ) ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘– )2 โ‰ค
                ๐‘–                                          ๐‘–                                                      ๐‘–

                                                                         (๐‘˜โˆ’1)
where the last step uses that |๐‘Ž ๐‘– | โ‰ค 1 and |1 โˆ’ 2๐‘ฅ ๐‘–                           | โ‰ค 1.
                                             (๐‘˜) 2
                                        ๐‘– (๐‘ฆ ๐‘– ) โ‰ค ๐›พ๐‘› โ‰ค 1/2 by Claim 3.1, the third term can be bounded as
                                    ร
      Finally, as |๐‘Ž ๐‘– | โ‰ค 1 and
                                                                  !2
                                          ร•                                      ร•
                                                           (๐‘˜)                                (๐‘˜)
                                                   ๐‘Ž 2๐‘– (๐‘ฆ ๐‘– )2        โ‰ค (1/2)          ๐‘Ž 2๐‘– (๐‘ฆ ๐‘– )2 .                                  (3.4)
                                           ๐‘–                                       ๐‘–

      Plugging these bounds and using that ๐œ† โ‰ค 1, (3.3) gives that,
                                           ร•            (๐‘˜) 
                            ๐”ผ ๐‘˜โˆ’1 ๐‘Œ๐‘˜2 โ‰ค 8๐œ‚   ๐‘Ž 2๐‘– ๐”ผ ๐‘˜โˆ’1 (๐‘ฆ ๐‘– )2 = โˆ’(8๐œ‚/๐œ†)๐”ผ[๐‘Œ๐‘˜ ].
                                  
                                                       ๐‘–

where the last equality uses (3.2).                                                                                                          
   By Claim 3.3, we can apply Theorem 2.1 with ๐›ผ = ๐œ†/8๐œ‚, provided that the conditions for
Theorem 2.1 hold. Indeed, ๐›ผ โ‰ค 1 holds as ๐œ† โ‰ค 1 and ๐œ‚ โ‰ฅ 1, and |๐‘Œ๐‘˜ | < 1 holds by Claim 3.1.
   Applying Theorem 2.1 now gives that Pr[๐‘๐‘‡ โˆ’ ๐‘0 โ‰ฅ ๐‘ก 0] โ‰ค exp(โˆ’๐œ†๐‘ก 0/8๐œ‚). As ๐‘0 = ๐œ‡ + ๐œ†๐‘ฃ, this
gives
                           Pr[๐‘๐‘‡ โˆ’ ๐œ‡ โˆ’ ๐œ†๐‘ฃ โ‰ฅ ๐‘ก 0] โ‰ค exp(โˆ’๐‘ก 0๐œ†/4๐œ‚).                         (3.5)
Setting ๐œ† = ๐‘ก 0/(๐‘ก 0 + 2๐‘ฃ) (note that this satisfies our assumption ๐œ† โ‰ค 1), so that ๐œ†๐‘ฃ = ๐‘ก 0 ๐‘ฃ/(๐‘ก 0 + 2๐‘ฃ) โ‰ค
๐‘ก 0/2, (3.5) implies that
                                   Pr[๐‘๐‘‡ โˆ’ ๐œ‡ โ‰ฅ 3๐‘ก 0/2] โ‰ค exp(โˆ’๐‘ก 0๐œ†/4๐œ‚).
Setting ๐‘ก = 3๐‘ก 0/2 and plugging the value of ๐œ† gives the desired result that
                                                                ๐‘ก 2 /(18๐œ‚)
                                                                                                   
                                        Pr[๐‘๐‘‡ โˆ’ ๐œ‡ โ‰ฅ ๐‘ก] โ‰ค exp โˆ’             .                                                                 
                                                               2(๐‘ฃ + ๐‘ก/3)

4      Applications
4.1     Rounding column-sparse LPs
Let ๐‘ฅ โˆˆ [0, 1]๐‘› be a fractional solution to ๐ด๐‘ฅ = ๐‘, where ๐ด โˆˆ โ„ ๐‘šร—๐‘› is an arbitrary ๐‘š ร— ๐‘› matrix.
Let ๐‘ก = max ๐‘—โˆˆ[๐‘›] ( ๐‘–โˆˆ[๐‘š] |๐‘Ž ๐‘–๐‘— |) be the maximum โ„“ 1 -norm of the columns of ๐ด. Beck and Fiala [12]
                   ร
gave a rounding method to find ๐‘‹ โˆˆ {0, 1} ๐‘› so that the maximum rounding error for any row
satisfies k๐ด๐‘‹ โˆ’ ๐‘ k โˆž = k๐ด(๐‘‹ โˆ’ ๐‘ฅ)k โˆž < ๐‘ก.

                             T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                                            14
                  O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

Beckโ€“Fiala rounding. We first recall the iterated rounding algorithm in [12]. The algorithm
starts with ๐‘ฅ 0 = ๐‘ฅ and proceeds in iterations. Consider some iteration ๐‘˜, and let ๐ด ๐‘˜ denote the
matrix ๐ด restricted to the alive coordinates. Call a row big if its โ„“ 1 -norm in ๐ด ๐‘˜ is strictly more
than ๐‘ก. The key point is that by an averaging argument, the number of big rows is strictly less
than ๐‘› ๐‘˜ as each column has โ„“ 1 -norm at most ๐‘ก and thus the total โ„“ 1 -norm of all entries ๐ด ๐‘˜ is at
most ๐‘ก๐‘› ๐‘˜ . The algorithm chooses ๐‘Š (๐‘˜) to be matrix consisting of the big rows of ๐ด ๐‘˜ and applies
the iterated rounding update.
    Let us now analyze the error. Fix some row ๐‘–. As long as this row ๐‘– is big, its rounding error
is 0 during the update steps. Consider the first iteration when this row is no longer big. Then,
no matter how the remaining alive variables are rounded in subsequent iterations, the error
incurred will be (strictly) less than its โ„“1 -norm, which is at most ๐‘ก.

Introducing slack. To apply Theorem 1.2, we can easily introduce ๐›ฟ-slack for any 0 โ‰ค ๐›ฟ < 1,
as follows. In iteration ๐‘˜, call a row big if its โ„“ 1 norm exceeds ๐‘ก/(1 โˆ’ ๐›ฟ), and by the argument
above the number of big rows is strictly less than ๐‘› ๐‘˜ (1 โˆ’ ๐›ฟ). Theorem 1.2 now directly gives the
following result.

Theorem 4.1. Given a matrix ๐ด with maximum โ„“1 -norm of any column at most ๐‘ก, and any ๐‘ฅ โˆˆ [0, 1]๐‘› ,
then for any 0 โ‰ค ๐›ฟ < 1 the algorithm returns ๐‘‹ โˆˆ {0, 1} ๐‘› such that k๐ด(๐‘‹ โˆ’ ๐‘ฅ)k โˆž โ‰ค ๐‘ก/(1 โˆ’ ๐›ฟ),
๐”ผ[๐‘‹๐‘– ] = ๐‘ฅ ๐‘– and ๐‘‹ satisfies ๐‘‚(1/๐›ฟ)-concentration.

   This implies the following useful corollary.

Corollary 4.2. Let ๐‘€ be a matrix, and ๐ด be some subset of rows of ๐‘€ so that the columns of ๐‘€ restricted
to ๐ด have โ„“1 -norm at most ๐‘ก. Setting ๐›ฟ = 1/2, the rounding error is at most 2๐‘ก for rows of ๐ด, while the
other rows of ๐‘€ have error similar to that as under randomized rounding.

Komlรณs setting. For a ๐‘š ร— ๐‘› matrix ๐ด, let ๐‘ก = max ๐‘—โˆˆ[๐‘›] ( ๐‘–โˆˆ[๐‘š] ๐‘Ž 2๐‘–๐‘— )1/2 denote the maximum
                                                                ร

โ„“ 2 -norm of the columns of ๐ด. The long-standing Komlรณs conjecture (together with a connection
between hereditary discrepancy and rounding due to [25]) states that any ๐‘ฅ โˆˆ [0, 1]๐‘› can be
rounded to ๐‘‹ โˆˆ {0, 1}p ๐‘› so that k๐ด(๐‘‹ โˆ’ ๐‘ฅ)k = ๐‘‚(๐‘ก). Currently, the best known bound on this
                                            โˆž
rounding error is ๐‘‚(๐‘ก log ๐‘š) [5, 8].
   An argument similar to that for Theorem 4.1 gives the following result in this setting.

Theorem 4.3. If ๐ด has maximum column โ„“ 2 -norm                           ๐‘›
             ๐‘›
                                           p ๐‘ก, then given any ๐‘ฅ โˆˆ [0, 1] , the algorithm returns
an ๐‘‹ โˆˆ {0, 1} satisfying k๐ด(๐‘‹ โˆ’ ๐‘ฅ)k โˆž = ๐‘‚(๐‘ก log ๐‘š) and the ๐‘‚(1)-concentration property.

Proof. We will apply Theorem 1.2 with ๐›ฟ = 1/2. During any iteration ๐‘˜, call row ๐‘– big if its
squared โ„“ 2 norm in ๐ด ๐‘˜ exceeds 2๐‘ก 2 . As the sum of squared entries of ๐ด ๐‘˜ is at most ๐‘ก 2 ๐‘› ๐‘˜ , the
number big rows is at most ๐‘› ๐‘˜ /2 and we set ๐‘Š (๐‘˜) to ๐ด ๐‘˜ restricted to the big rows.
    The ๐‘‚(1)-concentration follows directly from Theorem 1.2. To bound the error for rows of ๐ด,
we argue as follows. Fix a row ๐‘–. Clearly, row ๐‘– incurs zero error while it is big. Let ๐‘˜ be the first
iteration when row ๐‘– is not big, and condition on the randomness up to this point. Call an (alive)

                      T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                           15
                                                           N IKHIL BANSAL

coordinate ๐‘— large if |๐‘Ž ๐‘–๐‘— | โ‰ฅ ๐‘ก/ log ๐‘š, and let ๐ฟ ๐‘– denote the set of large coordinates in row ๐‘–. Let
                                         p

๐‘Žหœ ๐‘– denote the row ๐‘Ž ๐‘– with the coordinates in ๐ฟ removed. As ๐‘— ๐‘Ž 2๐‘–๐‘— โ‰ค 2๐‘ก we have |๐ฟ ๐‘– | โ‰ค 2 log ๐‘š
                                                                    ร

and so the rounding error due to the coordinates in ๐ฟ ๐‘– can be at most
                                ร•                                 ร•                        p
                                         |๐‘Ž ๐‘–๐‘— | โ‰ค |๐ฟ ๐‘– | 1/2 (           |๐‘Ž ๐‘–๐‘— | 2 )1/2 = ๐‘‚(๐‘ก log ๐‘š).
                                 ๐‘—โˆˆ๐ฟ ๐‘–                            ๐‘—โˆˆ๐ฟ ๐‘–


Applying the ๐‘‚(1)-concentration property of the rounded solution ๐‘‹, the error due to the entries
of ๐‘Žหœ ๐‘– satisfies

                                                                          ๐‘ 2 ๐‘ก 2 log ๐‘š
                 ๏ฃฎร•                                  ๏ฃน
                                     (๐‘˜)
                                             p
                       ๐‘Ž ๐‘–๐‘— (๐‘‹ ๐‘— โˆ’ ๐‘ฅ ๐‘— ) โ‰ฅ ๐‘๐‘ก log ๐‘š ๏ฃบ๏ฃบ = exp ยญโˆ’๐‘ 0 ร
                 ๏ฃฏ                                   ๏ฃบ       ยฉ                            ยช
              Pr ๏ฃฏ
                 ๏ฃฏ
                                                                     ๐‘—โˆ‰๐ฟ ๐‘Ž ๐‘–๐‘— + ๐‘€๐‘๐‘ก log ๐‘š
                                                                           2
                                                                                      p   ยฎ
                 ๏ฃฏ ๐‘—โˆ‰๐ฟ                               ๏ฃบ
                 ๏ฃฐ                                   ๏ฃป       ยซ                            ยฌ
for some fixed constant ๐‘ 0.
    As ๐‘—โˆ‰๐ฟ ๐‘Ž 2๐‘–๐‘— โ‰ค 2๐‘ก 2 and ๐‘€ โ‰ค ๐‘ก/ log ๐‘š, the right hand side is exp(โˆ’ฮฉ(๐‘๐‘ 0 log ๐‘š)). Choosing ๐‘
       ร                          p

large enough so that this is  1/๐‘š, the result follows by a union bound over the rows.        

4.2   Makespan minimization on unrelated machines
In the unrelated machine setting, there are ๐‘Ÿ jobs and ๐‘š machines, and each job ๐‘— โˆˆ [๐‘Ÿ] has
size ๐‘ ๐‘–๐‘— on a machine ๐‘– โˆˆ [๐‘š]. The goal is to assign all the jobs to machines to minimize the
maximum machine load.


LP formulation. The standard LP relaxation has fractional assignment variables ๐‘ฅ ๐‘–๐‘— โˆˆ [0, 1]
for ๐‘— โˆˆ [๐‘Ÿ] and ๐‘– โˆˆ [๐‘š]. Consider the smallest target makespan ๐‘‡ for which the following LP is
feasible.
                   ร•
                           ๐‘ ๐‘–๐‘— ๐‘ฅ ๐‘–๐‘—     โ‰ค๐‘‡               โˆ€๐‘– โˆˆ [๐‘š]               (load constraints)
                   ๐‘—โˆˆ[๐‘Ÿ]
                      ร•
                               ๐‘ฅ ๐‘–๐‘—      =1               โˆ€๐‘— โˆˆ [๐‘Ÿ]               (assignment constraints)
                     ๐‘–โˆˆ[๐‘š]
                               ๐‘ฅ ๐‘–๐‘—      =0               โˆ€๐‘–, ๐‘— such that ๐‘ ๐‘–๐‘— > ๐‘‡

The last constraint is valid for any integral solution, and so we can assume that ๐‘ max := max๐‘–๐‘— ๐‘ ๐‘–๐‘—
is at most ๐‘‡. In a celebrated result, [24] gave a rounding procedure that produces an integral
solution with makespan at most ๐‘‡ + ๐‘ max . We now sketch the iterated rounding based proof of
this result from [23].


Iterated rounding proof. As always, we start with ๐‘ฅ (0) = ๐‘ฅ and fix the variables that get
rounded to 0 or 1. Consider some iteration ๐‘˜. Let ๐‘› ๐‘˜ denote the number of fractional variables,

                      T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                  16
                  O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

and let ๐‘… ๐‘˜ denote the set of jobs that are still not integrally assigned to some machine. For a
machine ๐‘–, define the excess as               ร•
                                                            (๐‘˜)
                                     ๐‘’ ๐‘– :=          (1 โˆ’ ๐‘ฅ ๐‘–๐‘— ),                           (4.1)
                                                                (๐‘˜)
                                                       ๐‘—โˆˆ๐‘… ๐‘˜ : ๐‘ฅ ๐‘–๐‘— >0

and note that ๐‘’ ๐‘– is simply the maximum (fractional) number of extra jobs that can be possibly
assigned to ๐‘– if all the non-zero variables are rounded to 1. An elegant counting argument in
[23] shows that if ๐‘Š (๐‘˜) consists of load constraints for machines with ๐‘’ ๐‘– > 1, and assignment
constraints for jobs in ๐‘… ๐‘˜ , then rank(๐‘Š (๐‘˜) ) < ๐‘› ๐‘˜ .

Introducing slack. We now extend the argument of [23] to introduce some slack so that we
can apply Theorem 1.2. This will give the following result.

Theorem 4.4. Given any ๐›ฟ โˆˆ [0, 1/2), and a fractional solution ๐‘ฅ to the problem, there is a rounding
where the integral solution ๐‘‹ increases the load on any machine by ๐‘ max /(1 โˆ’ 2๐›ฟ), satisfies ๐”ผ[๐‘‹๐‘–๐‘— ] = ๐‘ฅ ๐‘–๐‘—
for all ๐‘–, ๐‘— and has ๐‘‚(1/๐›ฟ)-concentration.
                                                                                                                    (๐‘˜)
Proof. Consider some iteration ๐‘˜, and let ๐‘› ๐‘˜ denote the number of fractional variables ๐‘ฅ ๐‘–๐‘— โˆˆ (0, 1),
and let ๐‘… ๐‘˜ denote the jobs that are still not integrally assigned. Let ๐‘Ÿ ๐‘˜ = |๐‘… ๐‘˜ |. For a machine ๐‘–,
we define the excess ๐‘’ ๐‘– as in (4.1). Let ๐‘€ ๐‘˜ denote the set of machines with ๐‘’ ๐‘– > 1/(1 โˆ’ 2๐›ฟ).
    ๐‘Š (๐‘˜) will consist of load constraints for machines in ๐‘€ ๐‘˜ and assignment constraints for jobs
                                      (๐‘˜)                                                        (๐‘˜)
in ๐‘… ๐‘˜ . More precisely, the update ๐‘ฆ ๐‘–๐‘— will satisfy the following two conditions: (i) ๐‘— ๐‘ ๐‘–๐‘— ๐‘ฆ ๐‘–๐‘— = 0
                                                                                         ร
                                  (๐‘˜)
for all ๐‘– โˆˆ ๐‘€ ๐‘˜ and (ii) ๐‘– ๐‘ฆ ๐‘–๐‘— = 0, for all ๐‘— โˆˆ ๐‘… ๐‘˜ . We say that machine ๐‘– is protected in iteration ๐‘˜ if
                          ร

๐‘– โˆˆ ๐‘€ ๐‘˜ . For a protected machine, the fractional load does not change after an update. When a
machine ceases to be protected for the first time, the definition of excess ensures that its extra
load in subsequent iterations can be at most ๐‘ max /(1 โˆ’ 2๐›ฟ).
    It remains to show that rank(๐‘Š๐‘˜ ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ . As each job in ๐‘… ๐‘˜ contributes at least two
fractional variables to ๐‘› ๐‘˜ , we first note that

                                                          2๐‘Ÿ ๐‘˜ โ‰ค ๐‘› ๐‘˜ .                                                    (4.2)

Let ๐‘š ๐‘˜ = |๐‘€ ๐‘˜ |. Then we also have the following.
Claim 4.5. ๐‘š ๐‘˜ โ‰ค (1 โˆ’ 2๐›ฟ)(๐‘› ๐‘˜ โˆ’ ๐‘Ÿ ๐‘˜ ).

Proof. Clearly ๐‘š ๐‘˜ /(1 โˆ’ 2๐›ฟ) โ‰ค              ๐‘–โˆˆ๐‘€ ๐‘˜ ๐‘’ ๐‘– as each ๐‘– โˆˆ ๐‘€ ๐‘˜ has excess more than 1/(1 โˆ’ 2๐›ฟ). Next,
                                        ร

                 ร•             ร•         ร•                            ร•       ร•
                                                          (๐‘˜)                                  (๐‘˜)
                        ๐‘’๐‘– =                       (1 โˆ’ ๐‘ฅ ๐‘–๐‘— ) โ‰ค                        (1 โˆ’ ๐‘ฅ ๐‘–๐‘— ) = ๐‘› ๐‘˜ โˆ’ ๐‘Ÿ ๐‘˜ ,
                ๐‘–โˆˆ๐‘€ ๐‘˜          ๐‘–โˆˆ๐‘€ ๐‘˜ ๐‘—โˆˆ๐‘… :๐‘ฅ (๐‘˜) >0                    ๐‘–โˆˆ๐‘€ ๐‘—โˆˆ๐‘… :๐‘ฅ (๐‘˜) >0
                                         ๐‘˜    ๐‘–๐‘—                               ๐‘˜   ๐‘–๐‘—


where the first equality uses the definition of ๐‘’ ๐‘– and second uses the definition of ๐‘› ๐‘˜ and that
        (๐‘˜)
  ๐‘–โˆˆ๐‘€ ๐‘ฅ ๐‘–๐‘— = 1 for each job ๐‘— โˆˆ ๐‘… ๐‘˜ . Together this gives ๐‘š ๐‘˜ โ‰ค (1 โˆ’ 2๐›ฟ)(๐‘› ๐‘˜ โˆ’ ๐‘Ÿ ๐‘˜ ).
ร
                                                                                                

                         T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                               17
                                                N IKHIL BANSAL

   Multiplying (4.2) by ๐›ฟ and adding to the inequality in Claim 4.5 gives ๐‘š ๐‘˜ + ๐‘Ÿ ๐‘˜ โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ ,
which implies the result as rank(๐‘Š๐‘˜ ) โ‰ค ๐‘Ÿ ๐‘˜ + ๐‘š ๐‘˜ .                                           
Remark 4.6. Setting ๐›ฟ = 0 recovers the additive ๐‘ max result of [24]. Theorem 4.4 also generalizes
directly to ๐‘ž resources, where job ๐‘— has load vector ๐‘ ๐‘–๐‘— = (๐‘ ๐‘–๐‘— (1), . . . , ๐‘ ๐‘–๐‘— (๐‘ž)) on machine ๐‘–, and
the goal is to find an assignment ๐ด of jobs to machines to minimize max โ„Ž,๐‘– ( ๐‘—:๐ด(๐‘—)=๐‘– ๐‘ ๐‘–๐‘— (โ„Ž)). A
                                                                                          ร
direct modification of the proof above gives an additive ๐‘ž๐‘max /(1 โˆ’ 2๐›ฟ) error and the ๐‘‚(1/๐›ฟ)-
concentration property.

4.3   Minimum cost degree-bounded matroid basis
Instead of just the degree-bounded spanning tree problem, we consider the more general
matroid setting as all the arguments apply directly without additional work.

Minimum cost degree-bounded matroid-basis problem (DegMat). The input is a matroid ๐‘€
defined on elements ๐‘‰ with costs ๐‘ : ๐‘‰ โ†’ โ„+ and ๐‘š degree constraints specified by (๐‘† ๐‘— , ๐‘ ๐‘— ) for
๐‘— โˆˆ [๐‘š], where ๐‘† ๐‘— โŠ‚ ๐‘‰ and ๐‘ ๐‘— โˆˆ โ„ค+ . The goal is to find a minimum-cost base ๐ผ in ๐‘€ satisfying
the degree bounds, i. e., |๐ผ โˆฉ ๐‘† ๐‘— | โ‰ค ๐‘ ๐‘— for all ๐‘— โˆˆ [๐‘š]. The matroid ๐‘€ is given implicitly, by an
independence oracle (which given a query ๐ผ, returns whether ๐ผ is an independent set or not).

Iterated rounding algorithm. The natural LP formulation for the problem has the variables
๐‘ฅ ๐‘– โˆˆ [0, 1] for each element ๐‘– โˆˆ ๐‘‰ and the goal is to minimize the cost ๐‘– ๐‘ ๐‘– ๐‘ฅ ๐‘– , subject to the
                                                                        ร
following constraints.
                      ร•
                              ๐‘ฅ๐‘–   โ‰ค ๐‘Ÿ(๐‘‡)       โˆ€๐‘‡ โŠ‚ ๐‘‰                (rank constraints)
                      ๐‘–โˆˆ๐‘‡
                      ร•
                              ๐‘ฅ๐‘–   = ๐‘Ÿ(๐‘‰)                          (matroid base constraint)
                      ๐‘–โˆˆ๐‘‰
                      ร•
                              ๐‘ฅ๐‘–   โ‰ค ๐‘๐‘—      โˆ€๐‘— โˆˆ [๐‘š]               (degree constraints)
                      ๐‘–โˆˆ๐‘† ๐‘—

Here ๐‘Ÿ(ยท) is the rank function of ๐‘€.
    Given a feasible LP solution with cost ๐‘ โˆ— , [22, 10] gave an iterated rounding algorithm that
finds a solution with cost at most ๐‘ โˆ— and an additive degree violation of at most ๐‘ž โˆ’ 1. Here
๐‘ž = max๐‘– |{๐‘— : ๐‘– โˆˆ ๐‘† ๐‘— } is the maximum number of sets that contain any element ๐‘–. Note that
๐‘ž = 2 for the degree bounded spanning tree problem, as the elements here are edges and the
sets ๐‘† ๐‘— consist of edges incident to a vertex, so that each edges can lie in at most two such sets.
    We briefly sketch the argument in [22, 10]. The algorithm starts with ๐‘ฅ (0) = ๐‘ฅ and applies
iterated rounding as follows. Consider some iteration ๐‘˜. Let ๐ด ๐‘˜ denote the set of fractional
variables and let ๐‘› ๐‘˜ = |๐ด ๐‘˜ |. For a set ๐‘† ๐‘— , define the excess as
                                                       ร•
                                                                        (๐‘˜)
                                            ๐‘’ ๐‘— :=                (1 โˆ’ ๐‘ฅ ๐‘– ),                       (4.3)
                                                     ๐‘–โˆˆ๐ด ๐‘˜ โˆฉ๐‘† ๐‘—


                       T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                            18
                   O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

the maximum degree violation for ๐‘† ๐‘— even if all current fractional variables are rounded to 1.
    Let ๐ท ๐‘˜ be the set of indices ๐‘— of degree constraints with excess ๐‘’ ๐‘— โ‰ฅ ๐‘ž. The algorithm chooses
๐‘Š to consist of the degree constraints in ๐ท ๐‘˜ (call these protected constraints) and some
  (๐‘˜)

basis for the tight matroid rank constraints. An elegant counting argument then shows that
rank(๐‘Š๐‘˜ ) โ‰ค ๐‘› ๐‘˜ โˆ’ 1. The correctness follows since if a degree constraint is no longer protected,
then its excess is strictly below ๐‘ž, and by integrality of ๐‘ ๐‘— and the final rounded solution, the
degree violation can be at most ๐‘ž โˆ’ 1.

Introducing slack. We will extend the argument above in a straightforward way to introduce
some slack, and then apply Theorem 1.2 to obtain the following result.

Theorem 4.7. For any 0 < ๐›ฟ < 1, there is an algorithm for the DegMat problem that produces a basis
with additive degree violation strictly less than ๐‘ž/(1 โˆ’ 2๐›ฟ) and satisfies ๐‘‚(1/๐›ฟ)-concentration.

   Setting ๐›ฟ = 1/6 so that 2/(1 โˆ’ 2๐›ฟ) = 3, and noting that the degree violation is strictly less
than this bound, gives the following.

Corollary 4.8. For the minimum cost degree bounded spanning tree problem, given a fractional
solution ๐‘ฅ there is an algorithm to find a spanning tree with degree violation of plus two and satisfying
๐‘‚(1)-concentration.

     We now describe the argument. Consider iteration ๐‘˜. Let ๐ด ๐‘˜ be the set of fractional variables
and ๐‘› ๐‘˜ = |๐ด ๐‘˜ |. We need to specify how to choose ๐‘Š (๐‘˜) and show that rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ .
Let ๐ท ๐‘˜ denote the set of indices ๐‘— of degree constraints with excess ๐‘’ ๐‘— โ‰ฅ ๐‘ž/(1 โˆ’ 2๐›ฟ). Let โ„ฑ denote
the family of the tight matroid constraints that hold with equality, i. e., ๐‘–โˆˆ๐‘†โˆฉ๐ด ๐‘˜ ๐‘ฅ ๐‘– = ๐‘Ÿ ๐‘˜ (๐‘†), where
                                                                              ร
๐‘Ÿ ๐‘˜ is the rank function of the matroid ๐‘€ ๐‘˜ obtained from ๐‘€ by deleting elements with ๐‘ฅ ๐‘– = 0
and contracting those with ๐‘ฅ ๐‘– = 1. It is well-known [30] that there is a chain family of tight sets
๐’ž = {๐ถ1 , . . . , ๐ถโ„“ }, with ๐ถ1 โŠ‚ ๐ถ2 โŠ‚ ยท ยท ยท โŠ‚ ๐ถโ„“ , such that the rank constraint of every ๐‘† โˆˆ โ„ฑ lies in
the linear span of the constraints for sets in ๐’ž. Let ๐‘ ๐‘˜ = |๐’ž| and ๐‘‘ ๐‘˜ = |๐ท ๐‘˜ |. We set ๐‘Š (๐‘˜) to be the
degree constraints in ๐ท ๐‘˜ and the rank constraints in ๐’ž.

Claim 4.9. rank(๐‘Š (๐‘˜) ) โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ .

Proof. It suffices to show that ๐‘ ๐‘˜ + ๐‘‘ ๐‘˜ โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ . As each ๐‘ฅ ๐‘– is fractional and as the ranks
๐‘Ÿ ๐‘˜ (๐ถ) are integral, it follows that any two sets in chain family differ by at least two elements,
i. e., |๐ถ ๐‘–+1 \ ๐ถ ๐‘– | โ‰ฅ 2. This implies that ๐‘ ๐‘˜ โ‰ค ๐‘› ๐‘˜ /2. We also note that ๐‘Ÿ ๐‘˜ (๐ถ1 ) < ๐‘Ÿ ๐‘˜ (๐ถ2 ) ยท ยท ยท < and in
particular the rank ๐‘Ÿ(๐ถ ๐‘ ๐‘˜ ) of the largest set in ๐’ž is at least ๐‘ ๐‘˜ . This gives that ๐‘–โˆˆ๐ด ๐‘˜ ๐‘ฅ ๐‘– โ‰ฅ ๐‘ ๐‘˜ .
                                                                                         ร
     Next, as ๐‘’ ๐‘— โ‰ฅ ๐‘ž/(1 โˆ’ 2๐›ฟ) for each ๐‘— โˆˆ ๐ท ๐‘˜ , we have that ๐‘ž๐‘‘ ๐‘˜ โ‰ค (1 โˆ’ 2๐›ฟ) ๐‘—โˆˆ๐ท๐‘˜ ๐‘’ ๐‘— . Moreover, by
                                                                                   ร
definition of ๐‘’ ๐‘— ,              ร•        ร• ร•                    ร•
                                       ๐‘’๐‘— =                      (1 โˆ’ ๐‘ฅ ๐‘– ) =           ๐‘ž ๐‘– (1 โˆ’ ๐‘ฅ ๐‘– )
                               ๐‘—โˆˆ๐ท ๐‘˜          ๐‘—โˆˆ๐ท ๐‘˜ ๐‘–โˆˆ๐ด ๐‘˜ โˆฉ๐‘† ๐‘—                  ๐‘–โˆˆ๐ด ๐‘˜

where ๐‘ž ๐‘– = |{๐‘— : ๐‘– โˆˆ ๐‘† ๐‘— , ๐‘— โˆˆ ๐ท ๐‘˜ }| is the number of tight degree constraints in ๐ท ๐‘˜ that contain
element ๐‘–. As ๐‘ž ๐‘– โ‰ค ๐‘ž, the above is at most ๐‘ž ๐‘–โˆˆ๐ด ๐‘˜ (1 โˆ’ ๐‘ฅ ๐‘– ) โ‰ค ๐‘ž๐‘› ๐‘˜ โˆ’ ๐‘ž๐‘ ๐‘˜ , where we use that
                                                    ร

  ๐‘–โˆˆ๐ด ๐‘˜ ๐‘ฅ ๐‘– โ‰ฅ ๐‘ ๐‘˜ , and   ๐‘–โˆˆ๐ด ๐‘˜ 1 = |๐ด ๐‘˜ | = ๐‘› ๐‘˜ .
ร                       ร


                        T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                                  19
                                           N IKHIL BANSAL

    Together this gives that ๐‘‘ ๐‘˜ โ‰ค (1 โˆ’ 2๐›ฟ)(๐‘› ๐‘˜ โˆ’ ๐‘ ๐‘˜ ), and adding 2๐›ฟ times the inequality ๐‘ ๐‘˜ โ‰ค ๐‘› ๐‘˜ /2
to this gives that ๐‘‘ ๐‘˜ + ๐‘ ๐‘˜ โ‰ค (1 โˆ’ ๐›ฟ)๐‘› ๐‘˜ , which proves the desired claim.                            

   The degree violation property follows as before, since if a degree constraint is no longer
protected, then its excess is strictly below ๐‘ž/(1 โˆ’ 2๐›ฟ).
   Finally, we remark that as the underlying LP has exponential size, some care is needed in
implementing the rounding algorithm, in particular in maintaining the chain family and in
computing the step size of the walk. These issues are discussed in [11].


Acknowledgements
We thank the referees for their extremely thorough reading of the manuscript, catching various
errors, and for several useful comments and suggestions that significantly improved the
presentation.


References
 [1] Nima Anari and Shayan Oveis Gharan: Effective-resistance-reducing flows, spectrally
     thin trees, and asymmetric TSP. In Proc. 56th FOCS, pp. 20โ€“39. IEEE Comp. Soc., 2015.
     [doi:10.1109/FOCS.2015.11] 6

 [2] Sanjeev Arora, Alan M. Frieze, and Haim Kaplan: A new rounding procedure for the
     assignment problem with applications to dense graph arrangement problems. Math.
     Programming, 92(1):1โ€“36, 2002. [doi:10.1007/s101070100271] 3, 6

 [3] Arash Asadpour, Michel X. Goemans, Aleksander Madry, Shayan Oveis Gharan, and
     Amin Saberi: An ๐‘‚(log ๐‘›/log log ๐‘›)-approximation algorithm for the asymmetric traveling
     salesman problem. INFORMS, 65(4):1043โ€“1061, 2017. Preliminary version in SODAโ€™10.
     [doi:10.1287/opre.2017.1603] 2, 3, 6, 7

 [4] Arash Asadpour and Amin Saberi: An approximation algorithm for max-min fair allocation
     of indivisible goods. SIAM J. Comput., 39(7):2970โ€“2989, 2010. [doi:10.1137/080723491] 3

 [5] Wojciech Banaszczyk: Balancing vectors and Gaussian measures of ๐‘›-dimensional
     convex bodies. Random Struct. Algor., 12(4):351โ€“360, 1998. [doi:10.1002/(SICI)1098-
     2418(199807)12:4<351::AID-RSA3>3.0.CO;2-S] 15

 [6] Nikhil Bansal: Constructive algorithms for discrepancy minimization. In Proc. 51st FOCS,
     pp. 3โ€“10. IEEE Comp. Soc., 2010. [doi:10.1109/FOCS.2010.7] 12

 [7] Nikhil Bansal: On a generalization of iterated and randomized rounding. In Proc. 51st
     STOC, pp. 1125โ€“1135. ACM Press, 2019. [doi:10.1145/3313276.3316313] 1

                      T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                           20
                O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

 [8] Nikhil Bansal, Daniel Dadush, and Shashwat Garg: An algorithm for Komlรณs
     Conjecture matching Banaszczykโ€™s bound. SIAM J. Comput., 48(2):534โ€“553, 2019.
     [doi:10.1137/17M1126795] 8, 15
 [9] Nikhil Bansal and Shashwat Garg: Algorithmic discrepancy beyond partial coloring. In
     Proc. 49th STOC, pp. 914โ€“926. ACM Press, 2017. [doi:10.1145/3055399.3055490] 8, 10
[10] Nikhil Bansal, Rohit Khandekar, and Viswanath Nagarajan: Additive guarantees
     for degree-bounded directed network design. SIAM J. Comput., 39(4):1413โ€“1431, 2010.
     [doi:10.1137/080734340] 18
[11] Nikhil Bansal and Viswanath Nagarajan: Approximation-friendly discrepancy rounding.
     In Proc. 18th Integer Prog. Combinat. Optim. (IPCOโ€™16), volume 9682 of LNCS, pp. 375โ€“386.
     Springer, 2016. [doi:10.1007/978-3-319-33461-5_31] 3, 8, 20
[12] Jรณzsef Beck and Tibor Fiala: โ€œInteger-makingโ€ theorems. Discr. Appl. Math., 3(1):1โ€“8, 1981.
     [doi:10.1016/0166-218X(81)90022-6] 5, 14, 15
[13] Julius Borcea, Peter Brรคndรฉn, and Thomas M. Liggett: Negative dependence and the
     geometry of polynomials. J. AMS, 22(2):521โ€“567, 2009. [doi:10.1090/S0894-0347-08-00618-8]
     3
[14] Stรฉphane Boucheron, Gรกbor Lugosi, and Pascal Massart: Concentration In-
     equalities: A Nonasymptotic Theory of Independence. Oxford Univ. Press, 2013.
     [doi:10.1093/acprof:oso/9780199535255.001.0001] 3
[15] Chandra Chekuri, Jan Vondrรกk, and Rico Zenklusen: Dependent randomized rounding
     via exchange properties of combinatorial structures. In Proc. 51st FOCS, pp. 575โ€“584. IEEE
     Comp. Soc., 2010. [doi:10.1109/FOCS.2010.60] 2, 3, 6
[16] Chandra Chekuri, Jan Vondrรกk, and Rico Zenklusen: Multi-budgeted matchings and
     matroid intersection via dependent rounding. In Proc. 22nd Ann. ACMโ€“SIAM Symp. on
     Discrete Algorithms (SODAโ€™11), pp. 1080โ€“1097. SIAM, 2011. [doi:10.1137/1.9781611973082.82]
     2, 3, 6
[17] Devdatt P. Dubhashi and Alessandro Panconesi: Concentration of Measure for the Analysis of
     Randomized Algorithms. Cambridge Univ. Press, 2009. [doi:10.1017/CBO9780511581274] 3
[18] David A. Freedman: On tail probabilities for martingales. Ann. Probab., 3(1):100โ€“118, 1975.
     [doi:10.1214/aop/1176996452] 8
[19] Rajiv Gandhi, Samir Khuller, Srinivasan Parthasarathy, and Aravind Srinivasan: Depen-
     dent rounding and its applications to approximation algorithms. J. ACM, 53(3):324โ€“360,
     2006. [doi:10.1145/1147954.1147956] 2, 3, 6
[20] Nicholas J. A. Harvey and Neil Olver: Pipage rounding, pessimistic estimators and matrix
     concentration. In Proc. 25th Ann. ACMโ€“SIAM Symp. on Discrete Algorithms (SODAโ€™14), pp.
     926โ€“945. SIAM, 2014. [doi:10.1137/1.9781611973402.69] 3, 6

                    T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                     21
                                        N IKHIL BANSAL

[21] Richard M. Karp, Frank Thomson Leighton, Ronald L. Rivest, Clark D. Thompson,
     Umesh V. Vazirani, and Vijay V. Vazirani: Global wire routing in two-dimensional arrays.
     Algorithmica, 2:113โ€“129, 1987. [doi:10.1007/BF01840353] 5

[22] Tamรกs Kirรกly, Lap Chi Lau, and Mohit Singh: Degree bounded matroids and submodular
     flows. Combinatorica, 32(6):703โ€“720, 2012. [doi:10.1007/s00493-012-2760-6] 18

[23] Lap-Chi Lau, R. Ravi, and Mohit Singh: Iterative Methods in Combinatorial Optimization.
     Cambridge Univ. Press, 2011. [doi:10.1017/CBO9780511977152] 2, 16, 17

[24] Jan Karel Lenstra, David B. Shmoys, and ร‰va Tardos: Approximation algorithms
     for scheduling unrelated parallel machines. Math. Programming, 46:259โ€“271, 1990.
     [doi:10.1007/BF01585745] 6, 16, 18

[25] Lรกszlรณ Lovรกsz, Joel H. Spencer, and Katalin Vesztergombi: Discrepancy of set-systems
     and matrices. Europ. J. Combinat., 7(2):151โ€“160, 1986. [doi:10.1016/S0195-6698(86)80041-5]
     8, 15

[26] Shachar Lovett and Raghu Meka: Constructive discrepancy minimization by walking on
     the edges. SIAM J. Comput., 44(5):1573โ€“1582, 2015. [doi:10.1137/130929400] 3, 12

[27] Robin Pemantle: Towards a theory of negative dependence. J. Math. Phys., 41(3):1371โ€“1390,
     2000. [doi:10.1063/1.533200] 3

[28] Yuval Peres, Mohit Singh, and Nisheeth K. Vishnoi: Random walks in poly-
     topes and negative dependence. In Proc. 8th Innovations in Theoret. Comp. Sci.
     Conf. (ITCSโ€™17), pp. 50:1โ€“10. Schloss Dagstuhlโ€“Leibniz-Zentrum fuer Informatik, 2017.
     [doi:10.4230/LIPIcs.ITCS.2017.50] 3

[29] Thomas RothvoรŸ: The entropy rounding method in approximation algorithms. In Proc.
     23rd Ann. ACMโ€“SIAM Symp. on Discrete Algorithms (SODAโ€™12), pp. 356โ€“372. SIAM, 2012.
     [doi:10.1137/1.9781611973099.32] 3, 8

[30] Alexander Schrijver: Combinatorial Optimization: Polyhedra and Efficiency. Volume B.
     Springer, 2003. Springer. 19

[31] Mohit Singh and Lap Chi Lau: Approximating minimum bounded degree spanning trees
     to within one of optimal. J. ACM, 62(1):1:1โ€“19, 2015. [doi:10.1145/2629366] 6, 7

[32] Mohit Singh and Nisheeth K. Vishnoi: Entropy, optimization and counting. In Proc. 46th
     STOC, pp. 50โ€“59. ACM Press, 2014. [doi:10.1145/2591796.2591803] 3, 6

[33] Aravind Srinivasan: Distributions on level-sets with applications to approxima-
     tion algorithms.    In Proc. 42nd FOCS, pp. 588โ€“597. IEEE Comp. Soc., 2001.
     [doi:10.1109/SFCS.2001.959935] 3

                    T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                    22
                O N A G ENERALIZATION OF I TERATED AND R ANDOMIZED ROUNDING

[34] Ola Svensson, Jakub Tarnawski, and Lรกszlรณ A. Vรฉgh: A constant-factor approximation
     algorithm for the asymmetric traveling salesman problem. J. ACM, 67(6):37:1โ€“53, 2020.
     [doi:10.1145/3424306] 6

[35] Vijay V. Vazirani: Approximation Algorithms. Springer, 2001. ACM DL. 2

[36] David P. Williamson and David B. Shmoys: The Design of Approximation Algorithms. Cam-
     bridge Univ. Press, 2011. CUP. 2


AUTHOR

     Nikhil Bansal
     Department of Computer Science
     University of Michigan, Ann Arbor
     bansal gmail com
     https://bansal.engin.umich.edu/


ABOUT THE AUTHOR

     Nikhil Bansal is a professor in the Department of Computer Science at University of
        Michigan, Ann Arbor. He attended the Indian Institute of Technology, Mumbai
        for his B. Tech. degree, and received his Ph. D. from Carnegie Mellon University,
        Pittsburgh. He got fascinated by algorithms after taking an undergraduate class
        by Ajit A. Diwan, and has worked on various algorithmic questions since then.
        During his free time, he enjoys reading, hiking and doing Yoga.




                    T HEORY OF C OMPUTING, Volume 20 (6), 2024, pp. 1โ€“23                    23