1 The WorstCase Analysis of Algorithms
1.1 Comparing Incomparable Algorithms
Comparing different algorithms is hard. For almost any pair of algorithms and measure of algorithm performance, each algorithm will perform better than the other on some inputs. For example, the MergeSort algorithm takes time to sort length arrays, whether the input is already sorted or not, while the running time of the InsertionSort algorithm is on alreadysorted arrays but in general.^{1}^{1}1A quick reminder about asymptotic notation in the analysis of algorithms: for nonnegative realvalued functions and defined on the natural numbers, we write if there are positive constants and such that for all ; if there exist positive and with for all ; and if is both and .
The difficulty is not specific to running time analysis. In general, consider a computational problem and a performance measure Perf, with quantifying the “performance” of an algorithm for on an input . For example, could be the Traveling Salesman Problem (TSP),
could be a polynomialtime heuristic for the problem, and
could be the approximation ratio of —i.e., the ratio of the lengths of ’s output tour and an optimal tour—on the TSP instance .^{2}^{2}2In the Traveling Salesman Problem, the input is a complete undirected graph with a nonnegative cost for each edge , and the goal is to compute an ordering of the vertices that minimizes the length of the corresponding tour (with interpreted as ). Or could be the problem of testing primality, a randomized polynomialtime primalitytesting algorithm, andthe probability (over
’s internal randomness) that the algorithm correctly decides if the positive integer is prime. In general, when two algorithms have incomparable performance, how can we deem one of them “better than” the other?Worstcase analysis is a specific modeling choice in the analysis of algorithms, where the performance profile of an algorithm is summarized by its worst performance on any input of a given size (i.e., or , depending on the measure, where denotes the size of the input ). The “better” algorithm is then the one with superior worstcase performance. MergeSort, with its worstcase asymptotic running time of for length arrays, is better in this sense than InsertionSort, which has a worstcase running time of .
1.2 Benefits of WorstCase Analysis
While crude, worstcase analysis can be tremendously useful and, for several reasons, it has been the dominant paradigm for algorithm analysis in theoretical computer science.

A good worstcase guarantee is the bestcase scenario for an algorithm, certifying its generalpurpose utility and absolving its users from understanding which inputs are most relevant to their applications. Thus worstcase analysis is particularly well suited for “generalpurpose” algorithms that are expected to work well across a range of application domains (like the default sorting routine of a programming language).

Worstcase analysis is often more analytically tractable to carry out than its alternatives, such as averagecase analysis with respect to a probability distribution over inputs.

For a remarkable number of fundamental computational problems, there are algorithms with excellent worstcase performance guarantees. For example, the lion’s share of an undergraduate algorithms course comprises algorithms that run in linear or nearlinear time in the worst case.^{3}^{3}3Worstcase analysis is also the dominant paradigm in complexity theory, where it has led to the development of completeness and many other fundamental concepts.
1.3 Goals of the Analysis of Algorithms
Before critiquing the worstcase analysis approach, it’s worth taking a step back to clarify why we want rigorous methods to reason about algorithm performance. There are at least three possible goals:

Performance prediction.
The first goal is to explain or predict the empirical performance of algorithms. In some cases, the analyst acts as a natural scientist, taking an observed phenomenon like “the simplex method for linear programming is fast” as ground truth, and seeking a transparent mathematical model that explains it. In others, the analyst plays the role of an engineer, seeking a theory that gives accurate advice about whether or not an algorithm will perform well in an application of interest.

Identify optimal algorithms. The second goal is to rank different algorithms according to their performance, and ideally to single out one algorithm as “optimal.” At the very least, given two algorithms and for the same problem, a method for algorithmic analysis should offer an opinion about which one is “better.”

Develop new algorithms. The third goal is to provide a welldefined framework in which to brainstorm new algorithms. Once a measure of algorithm performance has been declared, the Pavlovian response of most computer scientists is to seek out new algorithms that improve on the stateoftheart with respect to this measure. The focusing effect catalyzed by such yardsticks should not be underestimated.
When proving or interpreting results in algorithm design and analysis, it’s important to be clear in one’s mind about which of these goals the work is trying to achieve.
What’s the report card for worstcase analysis with respect to these three goals?

Worstcase analysis gives an accurate performance prediction only for algorithms that exhibit little variation in performance across inputs of a given size. This is the case for many of the greatest hits of algorithms covered in an undergraduate course, including the running times of nearlineartime algorithms and of many canonical dynamic programming algorithms. For many more complex problems, however, the predictions of worstcase analysis are overly pessimistic (see Section 2).

For the second goal, worstcase analysis earns a middling grade—it gives good advice about which algorithm to use for some important problems (like many of those in an undergraduate course) and bad advice for others (see Section 2).

Worstcase analysis has served as a tremendously useful brainstorming organizer. For over a halfcentury, researchers striving to optimize worstcase algorithm performance have been led to thousands of new algorithms, many of them practically useful.
2 Famous Failures and the Need for Alternatives
For many problems a bit beyond the scope of an undergraduate course, the downside of worstcase analysis rears its ugly head. This section reviews four famous examples where worstcase analysis gives misleading or useless advice about how to solve a problem. These examples motivate the alternatives to worstcase analysis that are surveyed in Section 4 and described in detail in later chapters of the book.
2.1 The Simplex Method for Linear Programming
Perhaps the most famous failure of worstcase analysis concerns linear programming, the problem of optimizing a linear function subject to linear constraints (Figure 1). Dantzig proposed in the 1940s an algorithm for solving linear programs called the simplex method. The simplex method solves linear programs using greedy local search on the vertices of the solution set boundary, and variants of it remain in wide use to this day. The enduring appeal of the simplex method stems from its consistently superb performance in practice. Its running time typically scales modestly with the input size, and it routinely solves linear programs with millions of decision variables and constraints. This robust empirical performance suggested that the simplex method might well solve every linear program in a polynomial amount of time.
Klee and Minty (1972) showed by example that there are contrived linear programs that force the simplex method to run in time exponential in the number of decision variables (for all of the common “pivot rules” for choosing the next vertex). This illustrates the first potential pitfall of worstcase analysis: overly pessimistic performance predictions that cannot be taken at face value. The running time of the simplex method is polynomial for all practical purposes, despite the exponential prediction of worstcase analysis.
To add insult to injury, the first worstcase polynomialtime algorithm for linear programming, the ellipsoid method, is not competitive with the simplex method in practice.^{4}^{4}4Interiorpoint methods, developed five years later, lead to algorithms that both run in worstcase polynomial time and are competitive with the simplex method in practice. Taken at face value, worstcase analysis recommends the ellipsoid method over the empirically superior simplex method. One framework for narrowing the gap between these theoretical predictions and empirical observations is smoothed analysis, the subject of Part IV of this book; see Section 4.4 for an overview.
2.2 Clustering and Hard Optimization Problems
Clustering is a form of unsupervised learning (finding patterns in unlabeled data), where the informal goal is to partition a set of points into “coherent groups” (Figure
2). One popular way to coax this goal into a welldefined computational problem is to posit a numerical objective function over clusterings of the point set, and then seek the clustering with the best objective function value. For example, the goal could be to choose cluster centers to minimize the sum of the distances between points and their nearest centers (the median objective) or the sum of the squared such distances (the means objective). Almost all natural optimization problems that are defined over clusterings are hard.^{5}^{5}5Recall that a polynomialtime algorithm for an hard problem would yield a polynomialtime algorithm for every problem in —for every problem with efficiently verifiable solutions. Assuming the widelybelieved conjecture, every algorithm for an hard problem either returns an incorrect answer for some inputs or runs in superpolynomial time for some inputs (or both).In practice, clustering is not viewed as a particularly difficult problem. Lightweight clustering algorithms, like Lloyd’s algorithm for means and its variants, regularly return the intuitively “correct” clusterings of realworld point sets. How can we reconcile the worstcase intractability of clustering problems with the empirical success of relatively simple algorithms?^{6}^{6}6More generally, optimization problems are more likely to be hard than polynomialtime solvable. In many cases, even computing an approximately optimal solution is an hard problem. Whenever an efficient algorithm for such a problem performs better on realworld instances than (worstcase) complexity theory would suggest, there’s an opportunity for a refined and more accurate theoretical analysis.
One possible explanation is that clustering is hard only when it doesn’t matter. For example, if the difficult instances of an hard clustering problem look like a bunch of random unstructured points, who cares? The common use case for a clustering algorithm is for points that represent images, or documents, or proteins, or some other objects where a “meaningful clustering” is likely to exist. Could instances with a meaningful clustering be easier than worstcase instances? Part III of this book covers recent theoretical developments that support an affirmative answer; see Section 4.2 for an overview.
2.3 The Unreasonable Effectiveness of Machine Learning
The unreasonable effectiveness of modern machine learning algorithms has thrown the gauntlet down to researchers in algorithm analysis, and there is perhaps no other problem domain that calls out as loudly for a “beyond worstcase” approach.
To illustrate some of the challenges, consider a canonical supervised learning problem, where a learning algorithm is given a data set of objectlabel pairs and the goal is to produce a classifier that accurately predicts the label of asyetunseen objects (e.g., whether or not an image contains a cat). Over the past decade, aided by massive data sets and computational power, neural networks have achieved impressive levels of performance across a range of prediction tasks. Their empirical success flies in the face of conventional wisdom in multiple ways. First, there is a computational mystery: Neural network training usually boils down to fitting parameters (weights and biases) to minimize a nonconvex loss function, for example to minimize the number of classification errors the model makes on the training set. In the past such problems were written off as computationally intractable, but firstorder methods (i.e., variants of gradient descent) often converge quickly to a local optimum or even to a global optimum. Why?
Second, there is a statistical mystery: Modern neural networks are typically overparameterized, meaning that the number of parameters to fit is considerably larger than the size of the training data set. Overparameterized models are vulnerable to large generalization error (i.e., overfitting), since they can effectively memorize the training data without learning anything that helps classify asyetunseen data points. Nevertheless, stateoftheart neural networks generalize shockingly well—why? The answer likely hinges on special properties of both realworld data sets and the optimization algorithms used for neural network training (principally stochastic gradient descent). Part V of this book covers the stateoftheart explanations of these and other mysteries in the empirical performance of machine learning algorithms.
The beyond worstcase viewpoint can also contribute to machine learning by “stresstesting” the existing theory and providing a road map for more robust guarantees. While work in beyond worstcase analysis makes strong assumptions relative to the norm in theoretical computer science, these assumptions are usually weaker than the norm in statistical machine learning. Research in the latter field often resembles averagecase analysis, for example when data points are modeled as independent and identically distributed samples from some underlying structured distribution. The semirandom models described in Parts III and IV of this book serve as role models for blending adversarial and averagecase modeling to encourage the design of algorithms with robustly good performance.
2.4 Analysis of Online Algorithms
Online algorithms are algorithms that must process their input as it arrives over time. For example, consider the online paging problem, where there is a system with a small fast memory (the cache) and a big slow memory. Data is organized into blocks called pages, with up to different pages fitting in the cache at once. A page request results in either a cache hit (if the page is already in the cache) or a cache miss (if not). On a cache miss, the requested page must be brought into the cache. If the cache is already full, then some page in it must be evicted. A cache replacement policy is an algorithm for making these eviction decisions. Any systems textbook will recommend aspiring to the least recently used (LRU) policy, which evicts the page whose most recent reference is furthest in the past. The same textbook will explain why: Realworld page request sequences tend to exhibit locality of reference, meaning that recently requested pages are likely to be requested again soon. The LRU policy uses the recent past as a prediction for the near future. Empirically, it typically suffers fewer cache misses than competing policies like firstin firstout (FIFO).
Worstcase analysis, straightforwardly applied, provides no useful insights about the performance of different cache replacement policies. For every deterministic policy and cache size , there is a pathological page request sequence that triggers a page fault rate of 100%, even though the optimal clairvoyant replacement policy (known as Bélády’s furthestinthefuture algorithm) would have a page fault rate of at most (Exercise 1). This observation is troublesome both for its absurdly pessimistic performance prediction and for its failure to differentiate between competing replacement policies (like LRU vs. FIFO). One solution, described in Section 3, is to choose an appropriately finegrained parameterization of the input space and to assess and compare algorithms using parameterized guarantees.
2.5 The Cons of WorstCase Analysis
We should celebrate the fact that worstcase analysis works so well for so many fundamental computational problems, while at the same time recognizing that the cherrypicked successes highlighted in undergraduate algorithms can paint a potentially misleading picture about the range of its practical relevance. The preceding four examples highlight the chief weaknesses of the worstcase analysis framework.

Overly pessimistic performance predictions.
By design, worstcase analysis gives a pessimistic estimate of an algorithm’s empirical performance. In the preceding four examples, the gap between the two is embarrassingly large.

Can rank algorithms inaccurately. Overly pessimistic performance summaries can derail worstcase analysis from identifying the right algorithm to use in practice. In the online paging problem, it cannot distinguish between the FIFO and LRU policies; for linear programming, it implicitly suggests that the ellipsoid method is superior to the simplex method.

No data model. If worstcase analysis has an implicit model of data, then it’s the “Murphy’s Law” data model, where the instance to be solved is an adversarially selected function of the chosen algorithm.^{7}^{7}7Murphy’s Law: If anything can go wrong, it will. Outside of security applications, this algorithmdependent model of data is a rather paranoid and incoherent way to think about a computational problem.
In many applications, the algorithm of choice is superior precisely because of properties of data in the application domain, like meaningful solutions in clustering problems or locality of reference in online paging. Pure worstcase analysis provides no language for articulating such domainspecific properties of data. In this sense, the strength of worstcase analysis is also its weakness.
These drawbacks show the importance of alternatives to worstcase analysis, in the form of models that articulate properties of “relevant” inputs and algorithms that possess rigorous and meaningful algorithmic guarantees for inputs with these properties. Research in “beyond worstcase analysis” is a conversation between models and algorithms, with each informing the development of the other. It has both a scientific dimension, where the goal is to formulate transparent mathematical models that explain empirically observed phenomena about algorithm performance, and an engineering dimension, where the goals are to provide accurate guidance about which algorithm to use for a problem and to design new algorithms that perform particularly well on the relevant inputs.
3 Example: Parameterized Bounds in Online Paging
3.1 Parameterizing by Locality of Reference
Returning to the online paging example in Section 2.4, perhaps we shouldn’t be surprised that worstcase analysis fails to advocate LRU over FIFO. The empirical superiority of LRU is due to the special structure in realworld page request sequences (locality of reference), which is outside the language of pure worstcase analysis.
The key idea for obtaining meaningful performance guarantees for and comparisons between online paging algorithms is to parameterize page request sequences according to how much locality of reference they exhibit, and then prove parameterized worstcase guarantees. Refining worstcase analysis in this way leads to dramatically more informative results. Part I of the book describes many other applications of such finegrained input parameterizations; see Section 4.1 for an overview.
How should we measure locality in a page request sequence? One tried and true method is the working set model, which is parameterized by a function from the positive integers to that describes how many different page requests are possible in a window of a given length. Formally, we say that a page sequence conforms to if for every positive integer and every set of consecutive page requests in , there are requests for at most distinct pages. For example, the identity function imposes no restrictions on the page request sequence. A sequence can only conform to a sublinear function like or if it exhibits locality of reference.^{8}^{8}8The notation means the number , rounded up to the nearest integer. We can assume without loss of generality that , , and for all (Exercise 2).
We adopt as our performance measure the fault rate of an online algorithm on the page request sequence —the fraction of requests in on which suffers a page fault. We next state a performance guarantee for the fault rate of the LRU policy with a cache size of that is parameterized by a number . The parameter is defined below in (1); intuitively, it will be close to 0 for slowgrowing functions (i.e., functions that impose strong locality of reference) and close to 1 for functions that grow quickly (e.g., nearlinearly). This performance guarantee requires that the function is approximately concave in the sense that the number of inputs with value under (that is, ) is nondecreasing in (Figure 3).
1  2  3  3  4  4  4  5  
1  2  3  4  5  6  7  8 
Theorem 3.1 (Albers et al. (2005)).
With defined as in (1) below:

For every approximately concave function , cache size , and deterministic cache replacement policy, there are arbitrarily long page request sequences conforming to for which the policy’s page fault rate is at least .

For every approximately concave function , cache size , and page request sequence that conforms to , the page fault rate of the LRU policy is at most plus an additive term that goes to 0 with the sequence length.

There exists a choice of an approximately concave function , a cache size , and an arbitrarily long page request sequence that conforms to , such that the page fault rate of the FIFO policy is bounded away from .
Parts (a) and (b) prove the worstcase optimality of the LRU policy in a strong and finegrained sense, by and by. Part (c) differentiates LRU from FIFO, as the latter is suboptimal for some (in fact, many) choices of and .
The guarantees in Theorem 3.1 are so good that they are meaningful even when taken at face value—for strongly sublinear ’s, goes to 0 reasonably quickly with . The precise definition of for is
(1) 
where we abuse notation and interpret as the smallest value of such that . That is, denotes the smallest window length in which page requests for distinct pages might appear. As expected, for the function we have for all . (With no restriction on the input sequence, an adversary can force a 100% fault rate.) If , however, then scales with . Thus with a cache size of 10,000, the page fault rate is always at most 1%. If , then goes to 0 even faster with , roughly as .
3.2 Proof of Theorem 3.1
Part (a).
To prove the lower bound in part (a), fix an approximately concave function and a cache size . Fix a deterministic cache replacement policy .
We construct a page sequence that uses only distinct pages, so at any given time step there is exactly one page missing from the algorithm’s cache. (Assume that the algorithm begins with the first pages in its cache.) The sequence comprises blocks, where the th block consists of consecutive requests for the same page , where is the unique page missing from the algorithm ’s cache at the start of the block. (Recall that is the number of values of such that .) This sequence conforms to (Exercise 5).
By the choice of the ’s, incurs a page fault on the first request of a block, and not on any of the other (duplicate) requests of that block. Thus, algorithm suffers exactly page faults.
The length of the page request sequence is . Because , this sum equals which, using the definition of the ’s, equals . The algorithm’s page fault rate on this sequence matches the definition (1) of , as required. More generally, repeating the construction over and over again produces arbitrarily long page request sequences for which the algorithm has page fault rate .
Part (b).
To prove a matching upper bound for the LRU policy, fix an approximately concave function , a cache size , and a sequence that conforms to . Our fault rate target is a major clue to the proof (recall (1)): we should be looking to partition the sequence into blocks of length at least such that each block has at most faults. So consider groups of consecutive faults of the LRU policy on . Each such group defines a block, beginning with the first fault of the group, and ending with the page request that immediately precedes the beginning of the next group of faults (see Figure 4).
Claim: Consider a block other than the first or last. Consider the page requests in this block, together with the requests immediately before and after this block. These requests are for at least distinct pages.
The claim immediately implies that every block contains at least requests. Because there are faults per block, this shows that the page fault rate is at most (ignoring the vanishing additive error due to the first and last blocks), proving Theorem 3.1(b).
We proceed to the proof of the claim. Note that, in light of Theorem 3.1(c), it is essential that the proof uses properties of the LRU policy not shared by FIFO. Fix a block other than the first or last, and let be the page requested immediately prior to this block. This request could have been a page fault, or not (cf., Figure 4). In any case, is in the cache when this block begins. Consider the faults contained in the block, together with the th fault that occurs immediately after the block. We consider three cases.
First, if the faults occurred on distinct pages that are all different from , we have identified our distinct requests ( and the faults). For the second case, suppose that two of the faults were for the same page . How could this have happened? The page was brought into the cache after the first fault on , and was not evicted until there were requests for distinct pages other than after this page fault. This gives distinct page requests ( and the other distinct requests between the two faults on ). Finally, suppose that one of these faults was on the page . Because was requested just before the first of these faults, the LRU algorithm, subsequent to this request and prior to evicting , must have received requests for distinct pages other than . These requests, together with that for , give the desired distinct page requests.^{9}^{9}9The first two arguments apply also to the FIFO policy, but the third does not. Suppose was already in the cache when it was requested just prior to the block. Under FIFO, this request does not “reset ’s clock”; if it was originally brought into the cache long ago, FIFO might well evict on the block’s very first fault.
3.3 Discussion
Theorem 3.1 is an example of a “parameterized analysis” of an algorithm, where the performance guarantee is expressed as a function of parameters of the input other than its size. A parameter like measures the “easiness” of an input, much like matrix condition numbers in linear algebra. We will see many more examples of parameterized analyses later in the book.
There are several reasons to aspire toward parameterized performance guarantees.

A parameterized guarantee is a mathematically stronger statement, containing strictly more information about an algorithm’s performance than a worstcase guarantee parameterized solely by the input size.

A parameterized analysis can explain why an algorithm has good “realworld” performance even when its worstcase performance is poor. The approach is to first show that the algorithm performs well for “easy” values of the parameter (e.g., for and such that is close to 0), and then make a case that “realworld” instances are “easy” in this sense (e.g., have enough locality of reference to conform to a function with a small value of ). The latter argument can be made empirically (e.g., by computing the parameter on representative benchmarks) or mathematically (e.g., by positing a generative model and proving that it typically generates “easy” inputs). Results in smoothed analysis (see Section 4.4 and Part IV) typically follow this twostep approach.

A parameterized performance guarantee suggests when—for which inputs, and which application domains—a given algorithm should be used. (Namely, on the inputs where the performance of the algorithm is good!) Such advice is useful to someone who has no time or interest in developing their own algorithm from scratch, and merely wishes to be an educated client of existing algorithms.^{10}^{10}10For a familiar example, parameterizing the running time of graph algorithms by both the number of vertices and the number of edges provides guidance about which algorithms should be used for sparse graphs and which ones for dense graphs.

Finegrained performance characterizations can differentiate algorithms when worstcase analysis cannot (as with LRU vs. FIFO).

Formulating a good parameter often forces the analyst to articulate a form of structure in data, like the “amount of locality” in a page request sequence. Ideas for new algorithms that explicitly exploit such structure often follow soon thereafter.^{11}^{11}11The parameter showed up only in our analysis of the LRU policy; in other applications, the chosen parameter also guides the design of algorithms for the problem.
Useful parameters come in several flavors. The parameter in Theorem 3.1 is derived directly from the input to the problem, and later chapters contain many more examples of such inputbased parameters. It is also common to parameterize algorithm performance by properties of an optimal solution. In parameterized algorithms (Chapter 2), the most wellstudied such parameter is the size of an optimal solution. Another solutionbased parameterization, popular in machine learning applications, is by the “margin,” meaning the extent to which the optimal solution is particularly “pronounced”; see Exercise 9
for the canonical example of the analysis of the perceptron algorithm.
“Input size” is well defined for every computational problem, and this is one of the reasons why performance guarantees parameterized by input size are so ubiquitous. By contrast, the parameter used in Theorem 3.1 is specifically tailored to the online paging problem; in exchange, the performance guarantee is unusually accurate and meaningful. Alas, there are no silver bullets in parameterized analysis, or in algorithm analysis more generally, and the most enlightening analysis approach is often problemspecific. Worstcase analysis can inform the choice of an appropriate analysis framework for a problem by highlighting the problem’s most difficult (and often unrealistic) instances.
4 Overview of Book
This book has six parts, four on “core theory” and two on “applications.” Each of the following sections summarizes the chapters in one of the parts.
4.1 Refinements of WorstCase Analysis
Part I of the book hews closest to traditional worstcase analysis. No assumptions are imposed on the input; as in worstcase analysis, there is no commitment to a “model of data.” The innovative ideas in these chapters concern novel and problemspecific ways of expressing algorithm performance. Our online paging example (Section 3) falls squarely in this domain.
Chapter 2, by Fomin, Lokshtanov, Saurabh and Zehavi, provides an overview of the relatively mature field of parameterized algorithms. The goal here is to understand how the running time of algorithms and the complexity of computational problems depend on parameters other than the input size. For example, for which hard problems and parameters is “fixedparameter tractable” with respect to , meaning solvable in time for some function that is independent of the input size ? The field has developed a number of powerful approaches to designing fixedparameter tractable algorithms, as well as lower bound techniques for ruling out the existence of such algorithms (under appropriate complexity assumptions).
Chapter 3, by Barbay, searches for instanceoptimal algorithms—algorithms that for every input perform better than every other algorithm (up to a constant factor). Such an inputbyinput guarantee is essentially the strongest notion of optimality one could hope for. Remarkably, there are several fundamental problems, for example in lowdimensional computational geometry, that admit an instanceoptimal algorithm. Proofs of instance optimality involve inputbyinput matching upper and lower bounds, and this typically requires a very finegrained parameterization of the input space.
Chapter 4, by Roughgarden, concerns resource augmentation. This concept makes sense for problems that have a natural notion of a “resource,” with the performance of an algorithm improving as it is given more resources. Examples include the size of a cache (with larger caches leading to fewer faults), the capacity of a network (with highercapacity networks leading to less congestion), and the speed of a processor (with faster processors leading to earlier job completion times). A resource augmentation guarantee then states that the performance of an algorithm of interest is always close to that achieved by an allpowerful algorithm that is handicapped by slightly less resources.
4.2 Deterministic Models of Data
Part II of the book proposes deterministic models of data for several hard clustering and sparse recovery problems, which effectively posit conditions that are conceivably satisfied by “realworld” inputs. This work fits into the longstanding tradition of identifying “islands of tractability,” meaning polynomialtime solvable special cases of hard problems. 20thcentury research on tractable special cases focused primarily on syntactic and easilychecked restrictions (e.g., graph planarity or Horn satisfiability). The chapters in Part II and some of the related application chapters consider conditions that are not necessarily easy to check, but for which there is a plausible narrative about why “realworld instances” might satisfy them, at least approximately.
Chapter 5, by Makarychev and Makarychev, studies perturbation stability in several different computational problems. A perturbationstable instance satisfies a property that is effectively a uniqueness condition on steroids, stating that the optimal solution remains invariant to sufficiently small perturbations of the numbers in the input. The larger the perturbations that are tolerated, the stronger the condition on the instance and the easier the computational problem. Many problems have “stability thresholds,” an allowable perturbation size at which the complexity of the problem switches suddenly from hard to polynomialtime solvable. To the extent that we’re comfortable identifying “instances with a meaningful clustering” with perturbationstable instances, the positive results in this chapter give a precise sense in which clustering is hard only when it doesn’t matter (cf., Section 2.2). As a bonus, many of these positive results are achieved by algorithms that resemble popular approaches in practice, like singlelinkage clustering and local search.
Chapter 6, by Blum, proposes an alternative condition called approximation stability, stating that every solution with a nearoptimal objective function value closely resembles the optimal solution. That is, any solution that is structurally different from the optimal solution has significantly worse objective function value. This condition is particularly appropriate for problems like clustering, where the objective function is only means to an end and the real goal is to recover some type of “groundtruth” clustering. This chapter demonstrates that many hard problems become provably easier for approximationstable instances.
Chapter 7, by Price, provides a glimpse of the vast literature on sparse recovery, where the goal is to reverse engineer a “sparse” object from a small number of clues about it. This area is more strongly associated with applied mathematics than with theoretical computer science and algorithms, but there are compelling parallels between it and the topics of the preceding two chapters. For example, consider the canonical problem in compressive sensing, where the goal is to recover an unknown sparse signal
(a vector of length
) from a small number of linear measurements of it. If can be arbitrary, then the problem is hopeless unless . But many realworld signals have most of their mass concentrated on coordinates for small (and an appropriate basis), and the results surveyed in this chapter show that, for such “natural” signals, the problem can be solved efficiently even when is only modestly bigger than (and much smaller than ).4.3 SemiRandom Models
Part III of the book is about semirandom models—hybrids of worst and averagecase analysis in which nature and an adversary collaborate to produce an instance of a problem. For many problems, such hybrid frameworks are a “sweet spot” for algorithm analysis, with the worstcase dimension encouraging the design of robustly good algorithms and the averagecase dimension allowing for strong provable guarantees.
Chapter 8, by Roughgarden, sets the stage with a review of pure averagecase or distributional analysis, along with some of its killer applications and biggest weaknesses. Work in this area adopts a specific probability distribution over the inputs of a problem, and analyzes the expectation (or some other statistic) of the performance of an algorithm with respect to this distribution. One use of distributional analysis is to show that a generalpurpose algorithm has good performance on nonpathological inputs (e.g., deterministic QuickSort on randomly ordered arrays). One key drawback of distributional analysis is that it can encourage the design of algorithms that are brittle and overly tailored to the assumed input distribution. The semirandom models of the subsequent chapters are designed to ameliorate this issue.
Chapter 9, by Feige, introduces several planted models and their semirandom counterparts. For example, in the planted clique problem, a clique of size is planted in an otherwise uniformly random graph. How large does need to be, as a function of the number of vertices, before the planted clique can be recovered in polynomial time (with high probability)? In a semirandom version of a planted model, an adversary can modify the random input in a restricted way. For example, in the clique problem, an adversary might be allowed to remove edges not in the clique; such changes intuitively make the planted clique only “more obviously optimal,” but nevertheless can foil overly simplistic algorithms. One rule of thumb that emerges from this line of work, and also recurs in the next chapter, is that spectral algorithms tend to work well for planted models but the heavier machinery of semidefinite programming seems required for their semirandom counterparts. This chapter also investigates random and semirandom models for Boolean formulas, including refutation algorithms that certify that a given input formula is not satisfiable.
Chapter 10, by Moitra, drills down on a specific and extensivelystudied planted model, the stochastic block model. The vertices of a graph are partitioned into groups, and each potential edge of the graph is present independently with a probability that depends only on the groups that contain its endpoints. The algorithmic goal is to recover the groups from the (unlabeled) graph. One important special case is the planted bisection problem, where the vertices are split into two equalsize sets and and each edge is present independently with probability (if both endpoints are in the same group) or (otherwise). How big does the gap need to be before the planted bisection can be recovered, either statistically (i.e., with unbounded computational power) or with a polynomialtime algorithm? When and are sufficiently small, the relevant goal becomes partial recovery, meaning a proposed classification of the vertices with accuracy better than random guessing. In the semirandom version of the model, an adversary can remove edges crossing the bisection and add edges internal to each of the groups. For partial recovery, this semirandom version is provably more difficult than the original model.
Chapter 11, by Gupta and Singla, describes results for a number of online algorithms in randomorder models
. These are semirandom models in which an adversary decides on an input, and nature then presents this input to an online algorithm, one piece at a time and in random order. The canonical example here is the secretary problem, where an arbitrary finite set of numbers is presented to an algorithm in random order, and the goal is to design a stopping rule with the maximumpossible probability of stopping on the largest number of the sequence. Analogous randomorder models have proved useful for overcoming worstcase lower bounds for the online versions of a number of combinatorial optimization problems, including bin packing, facility location, and network design.
Chapter 12, by Seshadhri, is a survey of the field of selfimproving algorithms. The goal here is to design an algorithm that, when presented with a sequence of independent samples drawn from an unknown input distribution, quickly converges to the optimal algorithm for that distribution. For example, for many distributions over length arrays, there are sorting algorithms that make less than comparisons on average. Could there be a “master algorithm” that replicates the performance of a distributionoptimal sorter from only a limited number of samples from the distribution? This chapter gives a positive answer under the assumption that array entries are drawn independently (from possibly different distributions), along with analogous positive results for several fundamental problems in lowdimensional computational geometry.
4.4 Smoothed Analysis
Part IV of the book focuses on the semirandom models studied in smoothed analysis. In smoothed analysis, an adversary chooses an arbitrary input, and this input is then perturbed slightly by nature. The performance of an algorithm is then assessed by its worstcase expected performance, where the worst case is over the adversary’s input choice and the expectation is over the random perturbation. This analysis framework can be applied to any problem where “small random perturbations” make sense, including most problems with realvalued inputs. It can be applied to any measure of algorithm performance, but has proven most effective for running time analyses of algorithms that seem to run in superpolynomial time only on highly contrived inputs (like the simplex method). As with other semirandom models, smoothed analysis has the benefit of potentially escaping worstcase inputs, especially if they are “isolated” in the input space, while avoiding overfitting a solution to a specific distributional assumption. There is also a plausible narrative about why “realworld” inputs are captured by this framework: Whatever problem you’d like to solve, there are inevitable inaccuracies in its formulation from measurement errors, uncertainty, and so on.
Chapter 13, by Manthey, details several applications of smoothed analysis to the analysis of local search algorithms for combinatorial optimization problems. For example, the 2opt heuristic for the Traveling Salesman Problem is a local search algorithm that begins with an arbitrary tour and repeatedly improves the current solution using local moves that swap one pair of edges for another. In practice, local search algorithms like the 2opt heuristic almost always converge to a locally optimal solution in a small number of steps. Delicate constructions show that the 2opt heuristic, and many other local search algorithms, require an exponential number of steps to converge in the worst case. The results in this chapter use smoothed analysis to narrow the gap between worstcase analysis and empirically observed performance, establishing that many local search algorithms (including the 2opt heuristic) have polynomial smoothed complexity.
Chapter 14, by Dadush and Huiberts, surveys the first and most famous killer application of smoothed analysis, the SpielmanTeng analysis of the running time of the simplex method for linear programming. As discussed in Section 2.1, the running time of the simplex method is exponential in the worst case but almost always polynomial in practice. This chapter develops intuition for and outlines a proof of the fact that the simplex method, implemented with the shadow vertex pivot rule, has polynomial smoothed complexity with respect to small Gaussian perturbations of the entries of the constraint matrix. The chapter also shows how to interpret the successive shortestpath algorithm for the minimumcost maximumflow problem as an instantiation of this version of the simplex method.
Chapter 15, by Röglin, presents a third application of smoothed analysis, to the size of Pareto curves for multiobjective optimization problems. For example, consider the knapsack problem, where the input consists of items with values and sizes. One subset of the items dominates another if it has both a larger overall value and a smaller overall size, and the Pareto curve is defined as the set of undominated solutions. Pareto curves matter for algorithm design because many algorithms for multiobjective optimization problems (like the NemhauserUllmann knapsack algorithm) run in time polynomial in the size of the Pareto curve. For many problems, the Pareto curve has exponential size in the worst case but expected polynomial size in a smoothed analysis model. This chapter also presents a satisfyingly strong connection between smoothed polynomial complexity and worstcase pseudopolynomial complexity for linear binary optimization problems.
4.5 Applications in Machine Learning and Statistics
Part V of the book gives a number of examples of how the paradigms in Parts I–IV have been applied to problems in machine learning and statistics.
Chapter 16, by Balcan and Haghtalab, considers one of the most basic problems in supervised learning, that of learning an unknown halfspace. This problem is relatively easy in the noiseless case but becomes notoriously difficult in the worst case in the presence of adversarial noise. This chapter surveys a number of positive statistical and computational results for the problem under additional assumptions on the datagenerating distribution. One type of assumption imposes structure, such as logconcavity, on the marginal distribution over data points (i.e., ignoring their labels). A second type restricts the power of the adversary who introduces the noise, for example by allowing the adversary to mislabel a point only with a probability that is bounded away from .
Chapter 17, by Diakonikolas and Kane, provides an overview of recent progress in
robust highdimensional statistics
, where the goal is to design learning algorithms that have provable guarantees even when a small constant fraction of the data points has been adversarially corrupted. For example, consider the problem of estimating the meanof an unknown onedimensional Gaussian distribution
, where the input consists of samples from the distribution andadditional points defined by an adversary. The empirical mean of the data points is a good estimator of the true mean when there is no adversary, but adversarial outliers can distort the empirical mean arbitrarily. The median of the input points, however, remains a good estimator of the true mean even with a small fraction of corrupted data points. What about in more than one dimension? Among other results, this chapter describes a robust and efficiently computable estimator for learning the mean of a highdimensional Gaussian distribution.
Chapter 18, by Dasgupta and Kpotufe, investigates the twin topics of nearest neighbor search and classification. The former is algorithmic, and the goal is to design a data structure that enables fast nearest neighbor queries. The latter is statistical, and the goal is to understand the amount of data required before the nearest neighbor classifier enjoys provable accuracy guarantees. In both cases, novel parameterizations are the key to narrowing the gap between worstcase analysis and empirically observed performance—for search, a parameterization of the data set; for classification, of the allowable target functions.
Chapter 19, by Vijayaraghavan, is about computing a lowrank tensor decomposition. For example, given an
3tensor with entries
, the goal is to express as a linear combination of the minimumpossible number of rankone tensors (where a rankone tensor has entries of the form for some vectors , , and ). Efficient algorithms for this problem are an increasingly important tool in the design of learning algorithms; see also Chapters 20 and 21. This problem is hard in general. Jennrich’s algorithm solves in polynomial time the special case of the problem in which the three sets of vectors in the lowrank decomposition (the ’s, the ’s, and the ’s) are linearly independent. This result does not address the overcomplete regime, meaning tensors that have rank larger than dimension. (Unlike matrices, the rank of a tensor can be much larger than its smallest dimension.) For this regime, the chapter shows that a generalization of Jennrich’s algorithm has smoothed polynomial complexity.Chapter 20, by Ge and Moitra, concerns topic modeling
, which is a basic problem in unsupervised learning. The goal here is to process a large unlabeled corpus of documents and produce a list of meaningful topics and an assignment of each document to a mixture of topics. One approach to the problem is to reduce it to nonnegative matrix factorization (NMF)—the analog of a singular value decomposition of a matrix, with the additional constraint that both matrix factors are nonnegative. The NMF problem is hard in general, but this chapter proposes a condition on inputs, which is reasonable in a topic modeling context, under which the problem can be solved quickly in both theory and practice. The key assumption is that each topic has at least one “anchor word,” the presence of which strongly indicates that the document is at least partly about that topic.
Chapter 21, by Ma, studies the computational mystery outlined in Section 2.3: Why are local methods like stochastic gradient descent so effective in solving the nonconvex optimization problems that arise in supervised learning, such as computing the lossminimizing parameters for a given neural network architecture? This chapter surveys the quickly evolving stateoftheart on this topic, including a number of different restrictions on problems under which local methods have provable guarantees. For example, some natural problems have a nonconvex objective function that satisfies the “strict saddle condition,” which asserts that at every saddle point (i.e., a point with zero gradient that is neither a minimum nor a maximum) there is a direction with strictly negative curvature. Under this condition, variants of gradient descent provably converge to a local minimum (and, for some problems, a global minimum).
Chapter 22, by Hardt, tackles the statistical mystery discussed in Section 2.3: Why do overparameterized models like deep neural networks, which have many more parameters than training data points, so often generalize well in practice? While the jury is still out, this chapter surveys several of the leading explanations for this phenomenon, ranging from properties of optimization algorithms like stochastic gradient descent (including algorithmic stability and implicit regularization) to properties of data sets (such as marginbased guarantees).
Chapter 23, by G. Valiant and P. Valiant, presents two instance optimality results for distribution testing and learning. The chapter first considers the problem of learning a discretely supported distribution from independent samples, and describes an algorithm that learns the distribution nearly as accurately as would an optimal algorithm with advance knowledge of the true multiset of (unlabeled) probabilities of the distribution. This algorithm is instance optimal in the sense that, whatever the structure of the distribution, the learning algorithm will perform almost as well as an algorithm specifically tailored for that structure. The chapter then explores the problem of identity testing: given the description of a reference probability distribution, p, supported on a countable set, and sample access to an unknown distribution, q, the goal is to distinguish whether versus the case that p and q have total variation distance at least . This chapter presents a testing algorithm that has optimal sample complexity simultaneously for every distribution p and , up to constant factors.
4.6 Further Applications
The final part of the book, Part VI, gathers a number of additional applications of the ideas and techniques introduced in Parts I–III.
Chapter 24, by Karlin and Koutsoupias, surveys alternatives to worstcase analysis in the competitive analysis of online algorithms. There is a long tradition in online algorithms of exploring alternative analysis frameworks, and accordingly this chapter connects to many of the themes of Parts I–III.^{12}^{12}12Indeed, the title of this book is a riff on that of a paper in the competitive analysis of online algorithms (Koutsoupias and Papadimitriou, 2000). For example, the chapter includes results on deterministic models of data (e.g., the access graph model for restricting the allowable page request sequences) and semirandom models (e.g., the diffuse adversary model to blend worst and averagecase analysis).
Chapter 25, by Ganesh and Vardi, explores the mysteries posed by the empirical performance of Boolean satisfiability (SAT) solvers. Solvers based on backtracking algorithms such as the DavisPutnamLogemannLoveland (DPLL) algorithm frequently solve SAT instances with millions of variables and clauses in a reasonable amount of time. This chapter provides an introduction to conflictdriven clauselearning (CDCL) solvers and their connections to proof systems, followed by a highlevel overview of the stateoftheart parameterizations of SAT formulas, including inputbased parameters (such as parameters derived from the variableincidence graph of an instance) and outputbased parameters (such as the proof complexity in the proof system associated with CDCL solvers).
Chapter 26, by Chung, Mitzenmacher, and Vadhan, uses ideas from pseudorandomness to explain why simple hash functions work so well in practice. Welldesigned hash functions are practical proxies for random functions—simple enough to be efficiently implementable, but complex enough to “look random.” In the theoretical analysis of hash functions and their applications, one generally assumes that a hash function is chosen at random from a restricted family, such as a set of universal or wise independent functions for small . For some statistics, such as the expected number of collisions under a random hash function, small families of hash functions provably perform as well as completely random functions. For others, such as the expected insertion time in a hash table with linear probing, simple hash functions are provably worse than random functions (for worstcase data). The running theme of this chapter is that a little randomness in the data, in the form of a lower bound on the entropy of the (otherwise adversarial) datagenerating distribution, compensates for any missing randomness in a universal family of hash functions.
Chapter 27, by TalgamCohen, presents an application of the beyond worstcase viewpoint in algorithmic game theory, to
priorindependent auctions. For example, consider the problem of designing a singleitem auction, which solicits bids from bidders and then decides which bidder (if any) wins the item and what everybody pays. The traditional approach in economics to designing revenuemaximizing auctions is averagecase, meaning that the setup includes a commonly known distribution over each bidder’s willingness to pay for the item. An auction designer can then implement an auction that maximizes the expected revenue with respect to the assumed distributions (e.g., by setting a distributiondependent reserve price). As with many averagecase frameworks, this approach can lead to impractical solutions that are overly tailored to the assumed distributions. A semirandom variant of the model allows an adversary to pick its favorite distribution out of a rich class, from which nature chooses a random sample for each bidder. This chapter presents priorindependent auctions, both with and without a type of resource augmentation, that achieve nearoptimal expected revenue simultaneously across all distributions in the class.Chapter 28, by Roughgarden and Seshadhri, takes a beyond worstcase approach to the analysis of social networks. Most research in social network analysis revolves around a collection of competing generative models—probability distributions over graphs designed to replicate the most common features observed in such networks. The results in this chapter dispense with generative models and instead provide algorithmic or structural guarantees under deterministic combinatorial restrictions on a graph—that is, for restricted classes of graphs. The restrictions are motivated by the most uncontroversial properties of social and information networks, such as heavytailed degree distributions and strong triadic closure properties. Results for these graph classes effectively apply to all “plausible” generative models of social networks.
Chapter 29, by Balcan, reports on the emerging area of datadriven algorithm design. The idea here is to model the problem of selecting the bestinclass algorithm for a given application domain as an offline or online learning problem, in the spirit of the aforementioned work on selfimproving algorithms. For example, in the offline version of the problem, there is an unknown distribution over inputs, a class of allowable algorithms, and the goal is to identify from samples the algorithm in with the best expected performance with respect to . The distribution captures the details of the application domain, the samples correspond to benchmark instances representative of the domain, and the restriction to the class is a concession to the reality that it is often more practical to be an educated client of alreadyimplemented algorithms than to design a new algorithm from scratch. For many computational problems and algorithm classes , it is possible to learn an (almost) bestinclass algorithm from a modest number of representative instances.
Chapter 30, by Mitzenmacher and Vassilvitskii, is an introduction to algorithms with predictions. For example, in the online paging problem (Section 2.4), the LRU policy makes predictions about future page requests based on the recent past. If its predictions were perfect, the algorithm would be optimal. What if a good but imperfect predictor is available, such as one computed by a machine learning algorithm using past data? An ideal solution would be a generic online algorithm that, given a predictor as a “black box”: (i) is optimal when predictions are perfect; (ii) has gracefully degrading performance as the predictor error increases; and (iii) with an arbitrarily poor predictor, defaults to the optimal worstcase guarantee. This chapter investigates the extent to which properties (i)–(iii) can be achieved by predictoraugmented data structures and algorithms for several different problems.
5 Notes
This chapter is based in part on Roughgarden (2019).
The simplex method (Section 2.1) is described, for example, in Dantzig (1963); Khachiyan (1979) proved that the ellipsoid method solves linear programming problems in polynomial time; and the first polynomialtime interiorpoint method was developed by Karmarkar (1984). Lloyd’s algorithm for means (Section 2.2) appears in Lloyd (1962). The phrase “clustering is hard only when it doesn’t matter” (Section 2.2) is credited to Naftali Tishby by Daniely et al. (2012). The competitive analysis of online algorithms (Section 2.4) was pioneered by Sleator and Tarjan (1985). Bélády’s algorithm (Section 2.4) appears in Bélády (1967). The working set model in Section 3.1 was formulated by Denning (1968). Theorem 3.1 is due to Albers et al. (2005), as is Exercise 7. Exercise 8 is folklore. The result in Exercise 9 is due to Block (1962) and Novikoff (1962).
Acknowledgments
I thank Jérémy Barbay, Daniel Kane, and Salil Vadhan for helpful comments on a preliminary draft of this chapter.
References
 Albers et al. (2005) Albers, S., L. M. Favrholdt, and O. Giel (2005). On paging with locality of reference. Journal of Computer and System Sciences 70(2), 145–175.
 Bélády (1967) Bélády, L. A. (1967). A study of replacement algorithms for a virtual storage computer. IBM Systems Journal 5(2), 78–101.
 Block (1962) Block, H. D. (1962). The perceptron: A model for brain functioning. Reviews of Modern Physics 34, 123–135.
 Daniely et al. (2012) Daniely, A., N. Linial, and M. Saks (2012). Clustering is difficult only when it does not matter. arXiv:1205.4891.
 Dantzig (1963) Dantzig, G. B. (1963). Linear Programming and Extensions. Princeton University Press.
 Denning (1968) Denning, P. J. (1968, May). The working set model for program behavior. Commuications of the ACM 11(5), 323–333.
 Karmarkar (1984) Karmarkar, N. (1984). A new polynomialtime algorithm for linear programming. Combinatorica 4, 373–395.
 Khachiyan (1979) Khachiyan, L. G. (1979). A polynomial algorithm in linear programming. Soviet Mathematics Doklady 20(1), 191–194.
 Klee and Minty (1972) Klee, V. and G. J. Minty (1972). How Good is the Simplex Algorithm? In O. Shisha (Ed.), Inequalities III, pp. 159–175. New York: Academic Press Inc.
 Koutsoupias and Papadimitriou (2000) Koutsoupias, E. and C. H. Papadimitriou (2000). Beyond competitive analysis. SIAM Journal on Computing 30(1), 300–317.
 Lloyd (1962) Lloyd, S. P. (1962). Least squares quantization in PCM. IEEE Transactions on Information Theory 28(2), 129–136.
 Novikoff (1962) Novikoff, A. (1962). On convergence proofs for perceptrons. In Proceedings of the Symposium on Mathematical Theory of Automata, Volume 12, pp. 615–622.
 Roughgarden (2019) Roughgarden, T. (2019). Beyond worstcase analysis. Communications of the ACM 62(3), 88–96.
 Roughgarden (2020) Roughgarden, T. (Ed.) (2020). Beyond the WorstCase Analysis of Algorithms. Cambridge University Press.
 Sleator and Tarjan (1985) Sleator, D. D. and R. E. Tarjan (1985). Amortized efficiency of list update and paging rules. Commincations of the ACM 28(2), 202–208.
Exercises

Prove that for every deterministic cache replacement policy and cache size , there is an adversarial page request sequence such that the policy faults on every request, and such that an optimal clairvoyant policy would fault on at most a fraction of the requests.
[Hint: use only distinct pages, and the fact that the optimal policy always evicts the page that will be requested furthest in the future.]

Let be a function of the type described in Section 3, with denoting the maximum allowable number of distinct page requests in any window of length .

Prove that there is a nondecreasing function with and for all such that a page request sequence conforms to if and only if it conforms to .

Prove that parts (a) and (b) of Theorem 3.1 hold trivially if .

Prove that the page request sequence constructed in the proof of Theorem 3.1(a) conforms to the given approximately concave function .
1 2 3 3 4 4 5 5 1 2 3 4 5 6 7 Figure 5: Function used to construct a bad page sequence for FIFO (Exercise 6). 
Prove Theorem 3.1(c).
[Hint: Many different choices of and work. For example, take , a set of 5 pages, the function shown in Figure 5, and a page request sequence consisting of an arbitrarily large number of identical blocks of the eight page requests 10203040.]

Prove the following analog of Theorem 3.1(b) for the FIFO replacement policy: for every and approximately concave function with , , and for all , the page fault rate of the FIFO policy on every request sequence that conforms to is at most
(2) 
An instance of the knapsack problem consists of items with nonnegative values and sizes , and a knapsack capacity . The goal is to compute a subset of items that fits in the knapsack (i.e., with ) and, subject to this, has the maximum total value .
One simple greedy algorithm for the problem reindexes the items in nonincreasing order of density and then returns the largest prefix of items that fits in the knapsack (i.e., with ). Parameterize a knapsack instance by the ratio of the largest size of an item and the knapsack capacity, and prove a parameterized guarantee for the greedy algorithm: the total value of its solution is at least times that of an optimal solution.

The perceptron algorithm is one of the most classical machine learning algorithms (Figure 6). The input to the algorithm is points in , with a label for each point . The goal is to compute a separating hyperplane
: a hyperplane with all of the positivelylabeled points on one side, and all of the negatively labeled points on the other. Assume that there exists a separating hyperplane, and moreover that some such hyperplane passes through the origin.
^{14}^{14}14The second assumption is without loss of generality, as it can be enforced by adding an extra “dummy coordinate” to the data points. We are then free to scale each data point so that —this does not change which side of a hyperplane is on.Parameterize the input by its margin , defined as
where ranges over the unit normal vectors of all separating hyperplanes. Let attain the maximum. Geometrically, the parameter is the smallest cosine of an angle defined by a point and the normal vector .

Prove that the squared norm of grows slowly with the number of iterations : for every .

Prove that the projection of onto grows significantly with every iteration: for every .

Conclude that the iteration count never exceeds .
Comments
There are no comments yet.