{
  "id": "approximation-algorithm",
  "title": "Approximation Algorithm",
  "lang": "en",
  "source_lang": "en",
  "summary": "An approximation algorithm efficiently finds feasible solutions to optimization problems with mathematically proven bounds on their distance from optimality.",
  "category": "computer-science",
  "subcategory": "algorithms-theory",
  "type": "concept",
  "aliases": [
    "Approximation algorithms",
    "provably approximate optimization algorithm"
  ],
  "generated_at": "2026-11-03",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/approximation-algorithm/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/approximation-algorithm/",
      "json": "https://aiwiki.page/zh/approximation-algorithm.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "An approximation algorithm is an algorithm that finds a feasible solution to an optimization problem together with a provable guarantee on the solution’s quality relative to an optimal solution. The term usually refers to polynomial-time algorithms for problems whose exact solution is computationally difficult. Rather than requiring the best possible answer, approximation algorithms establish how much quality must be sacrificed to obtain a solution efficiently.\n\nThe subject is especially important in combinatorial optimization, where applications involve choosing routes, assigning jobs, locating facilities, or selecting subsets under constraints. Many such problems are NP-hard, making a generally efficient exact algorithm unlikely under standard complexity assumptions. Approximation algorithms provide a systematic alternative: their performance is supported by mathematical analysis rather than only by successful experiments.\n\nThe field connects algorithm design with computational complexity, geometry, probability, and continuous optimization. Its central questions concern both possibilities and limits: which guarantees can be achieved, what computational resources they require, and whether a better guarantee would imply an unexpected breakthrough such as a polynomial-time algorithm for every problem in NP."
    },
    {
      "heading": "Formal guarantees and their interpretation",
      "text": "An optimization problem specifies a collection of instances, the feasible solutions for each instance, an objective function, and whether that objective is to be minimized or maximized. Write \\(\\operatorname{OPT}(I)\\) for the optimal objective value of instance \\(I\\), and \\(A(I)\\) for the value returned by an algorithm. Under the common convention that an approximation ratio is at least one, a \\(\\rho\\)-approximation for minimization satisfies\n\n\\[A(I)\\leq \\rho\\,\\operatorname{OPT}(I),\\]\n\nwhereas a \\(\\rho\\)-approximation for maximization satisfies\n\n\\[A(I)\\geq \\frac{\\operatorname{OPT}(I)}{\\rho}.\\]\n\nFeasibility is part of the guarantee: a low-cost solution that violates a required constraint is not ordinarily an approximation solution.\n\nFor maximization, another widespread convention reports the fraction of the optimum obtained. Under this convention an \\(\\alpha\\)-approximation, with \\(0<\\alpha\\leq1\\), guarantees \\(A(I)\\geq\\alpha\\operatorname{OPT}(I)\\). Thus “a factor-two approximation” and “a one-half approximation” can describe the same maximization guarantee. Statements about approximation must specify or make clear which convention is in use. The factor may be a constant, such as \\(2\\), or a function of instance size, such as \\(O(\\log n)\\).\n\nMultiplicative guarantees are most natural for nonnegative objectives. If a minimization instance has optimum zero, a finite multiplicative guarantee requires an output of value zero as well. Negative objective values require different definitions or restrictions. Moreover, adding a constant to every objective value preserves which solution is optimal but can radically change multiplicative approximability. Approximation quality is therefore a property of an objective’s numerical formulation, not merely of its ranking of solutions.\n\nAn additive guarantee instead bounds the absolute error, for example \\(A(I)\\leq\\operatorname{OPT}(I)+b\\) in minimization. Some algorithms have mixed bounds of the form \\(\\alpha\\operatorname{OPT}(I)+b\\). Others have asymptotic guarantees, in which a stated ratio applies as the optimum becomes large. These distinctions are particularly important when objective values are small integers.\n\nThe usual efficiency requirement is polynomial time complexity in the input’s encoded length, including the bits used to represent weights and capacities. The guarantee is normally worst-case over all valid instances, not an average over a convenient dataset. A randomized algorithm may instead guarantee an expected objective value or a bound that holds with specified probability. Such statements must distinguish randomness used by the algorithm from randomness in the input. A deterministic guarantee for every instance, an expectation guarantee for every instance, and an empirical average are different kinds of evidence."
    },
    {
      "heading": "Historical development and the complexity perspective",
      "text": "Approximation methods predate the modern complexity classification of optimization problems. In scheduling, Ronald Graham’s 1966 analysis of list scheduling gave a particularly influential worst-case guarantee. When jobs are assigned one by one to the currently least-loaded of \\(m\\) identical machines, the resulting maximum load is at most \\(2-1/m\\) times optimal. Graham’s later analysis of ordering jobs by decreasing processing time yielded the stronger bound \\(4/3-1/(3m)\\). These results showed that simple procedures could support precise, instance-independent performance claims.\n\nThe development of NP-completeness in the early 1970s changed the motivation for such work. The results of Stephen Cook and Richard Karp identified broad families of problems whose exact solution was linked to the P versus NP problem. Approximation became a way to distinguish difficult optimization problems more finely. Exact NP-hardness does not determine whether a problem admits a near-optimal solution efficiently: some NP-hard problems permit arbitrarily accurate polynomial-time approximation, while others resist every fixed constant factor.\n\nSeveral foundational results appeared during the 1970s. David Johnson’s 1974 work helped establish systematic approximation analysis for combinatorial problems, including logarithmic guarantees for covering. Sartaj Sahni developed early approximation schemes for scheduling and knapsack-type problems. Oscar Ibarra and Chul Kim’s 1975 work gave fully polynomial approximation schemes for knapsack and subset sum. In 1976, Nicos Christofides introduced a \\(3/2\\)-approximation for the metric traveling salesperson problem. These results demonstrated markedly different levels of approximability within the same broad class of computationally difficult problems.\n\nDuring the 1980s and 1990s, mathematical programming became increasingly central. Relaxations allowed discrete problems to be compared with continuous lower or upper bounds. Primal–dual methods, randomized rounding, and semidefinite optimization supplied techniques that applied across many problems. In 1995, Michel Goemans and David Williamson presented their influential semidefinite approximation algorithm for maximum cut, obtaining a guarantee of approximately \\(0.87856\\) of the optimum.\n\nA parallel development concerned impossibility results. The theory of probabilistically checkable proofs, developed through work in the early 1990s, enabled strong statements about the difficulty of obtaining approximately optimal solutions. Later results connected exact approximation thresholds with conjectures about constraint satisfaction. The field consequently acquired two complementary sides: constructive algorithms proving that a ratio is achievable, and complexity reductions showing that improving it would require overcoming a stated computational assumption. A bound is especially informative when these two sides meet, because it identifies the best possible polynomial-time guarantee under that assumption."
    },
    {
      "heading": "Greedy selection, charging arguments, and local search",
      "text": "A greedy algorithm makes a sequence of locally attractive choices. Greedy methods are not automatically approximation algorithms: their quality must be proved for the particular problem and selection rule. A typical proof assigns the algorithm’s cost to objects that also constrain every optimal solution. This technique, often called a charging argument, converts a sequence of local decisions into a global comparison with the optimum.\n\nThe weighted set cover problem is a standard example. Given a universe of \\(n\\) elements and sets with nonnegative costs, the task is to cover every element at minimum total cost. The greedy rule chooses the set minimizing its cost divided by the number of newly covered elements. If \\(r\\) elements remain uncovered and an optimal cover costs \\(C\\), at least one set in that cover has cost per newly covered element at most \\(C/r\\). Charging each newly covered element the cost per element of its selected set yields\n\n\\[\\operatorname{cost}(\\text{greedy})\\leq H_n C, \\qquad H_n=\\sum_{i=1}^{n}\\frac1i\\leq 1+\\ln n.\\]\n\nThe logarithmic guarantee follows from the declining number of elements still available to share the remaining cost.\n\nA different elementary argument gives a factor-two approximation for unweighted vertex cover. Construct any maximal matching, meaning a set of disjoint edges to which no additional edge can be added. Output both endpoints of every matched edge. This is a vertex cover: an uncovered edge would have two unmatched endpoints and could have been added to the matching. Every feasible vertex cover must include at least one endpoint of each matched edge, so the matching size is a lower bound on the optimum, while the algorithm selects twice that many vertices. A maximum matching is unnecessary; maximality alone suffices.\n\nLocal search starts with a feasible solution and repeatedly applies an improving move from a defined neighborhood. A local optimum need not be globally optimal, but carefully chosen neighborhoods can imply approximation guarantees. In metric \\(k\\)-median, for example, exchanging one open facility for one closed facility yields a factor-five guarantee at an exact local optimum. Allowing exchanges involving up to \\(p\\) facilities gives a bound of \\(3+2/p\\). The proof compares local facilities with optimal ones and combines inequalities showing that specified exchanges cannot improve the current solution.\n\nLocal-search analysis must address running time as well as solution quality. With binary-encoded costs, a sequence of strictly improving moves can be extremely long. Algorithms therefore often require each accepted move to improve the objective by a sufficient relative amount. This can ensure polynomially many iterations, usually at the price of a small additional approximation loss. The neighborhood definition and stopping criterion are consequently parts of the theorem, not implementation details that can be omitted from it."
    },
    {
      "heading": "Linear relaxations, rounding, and primal–dual methods",
      "text": "Many discrete optimization problems can be written as integer programs, with variables restricted to integer values. Dropping those restrictions produces a linear programming relaxation. For minimization, every integral feasible solution is also feasible for the relaxation, so the fractional optimum is a lower bound on the true optimum. For maximization, it supplies an upper bound. The algorithmic task is then to convert the fractional solution into a legal discrete one without losing too much objective value.\n\nWeighted vertex cover illustrates deterministic rounding. Assign a variable \\(x_v\\) to each vertex \\(v\\), interpreted as whether that vertex is selected. The relaxation is\n\n\\[\\min \\sum_v w_vx_v \\quad\\text{subject to}\\quad x_u+x_v\\geq1\\ \\text{for every edge }uv,\\qquad x_v\\geq0.\\]\n\nSelect every vertex with \\(x_v\\geq1/2\\). Each edge has at least one selected endpoint, and\n\n\\[\\sum_{v:x_v\\geq1/2}w_v \\leq2\\sum_v w_vx_v \\leq2\\operatorname{OPT}.\\]\n\nThe threshold simultaneously establishes feasibility and controls cost.\n\nThe integrality gap measures the separation between an integer optimum and its relaxation. For a minimization problem it is the supremum, over applicable instances, of the integer optimum divided by the fractional optimum. A large gap limits what can be proved solely by comparing an output with that relaxation’s value. It does not by itself rule out a better approximation based on a stronger formulation or an unrelated argument. Understanding whether a limitation belongs to a particular relaxation or to the computational problem is a central distinction.\n\nThe primal–dual method exploits the dual linear program, whose feasible solutions certify bounds on the primal optimum. In weighted vertex cover, the dual places nonnegative values \\(y_e\\) on edges, subject to\n\n\\[\\sum_{e\\ni v}y_e\\leq w_v\\]\n\nat every vertex. Raising dual edge values and selecting vertices whose constraints become tight can produce a feasible cover. Its cost is bounded by twice the total dual value because each edge contributes to at most two selected endpoints. Weak duality then relates that total to the optimal cover cost.\n\nMore elaborate rounding methods preserve correlations, construct combinatorial structures, or repair violations after an initial selection. Facility location, network design, and scheduling have all benefited from such approaches. Relaxations are valuable not only because they are solvable, but because their constraints reveal the structure that a rounding algorithm must preserve. A fractional solution may identify which choices are cheap, which requirements are difficult to satisfy, and where coordinated decisions are necessary."
    },
    {
      "heading": "Randomization and semidefinite optimization",
      "text": "Randomness can simplify both construction and analysis. For maximum cut, a graph’s vertices must be divided into two parts to maximize the weight of edges crossing between them. If each vertex independently chooses either part with probability \\(1/2\\), every edge crosses with probability \\(1/2\\). The expected cut weight is therefore half the total edge weight and at least half the optimum. This establishes a one-half-of-optimum guarantee without knowing an optimal cut.\n\nThe same argument can be derandomized by the method of conditional expectations. Vertices are assigned one at a time, each time choosing a side that does not decrease the conditional expected final weight. Because the initial expectation is half the total edge weight, the completed cut achieves at least that value deterministically. This illustrates an important principle: randomness may expose an averaging argument even when a deterministic algorithm with the same bound is available.\n\nRandomized rounding applies probabilistic choices to a fractional relaxation. In set cover, a set with fractional value \\(x_j\\) can be selected with probability \\(\\min\\{1,cx_j\\ln n\\}\\), for a suitable constant \\(c\\). For an element whose covering variables sum to at least one, independence bounds its probability of remaining uncovered by approximately \\(n^{-c}\\). A union bound controls the probability that any element is uncovered, while linearity of expectation bounds the selected cost. A complete algorithm can use repetition or an explicit repair step; an expectation calculation alone does not establish that every sampled solution is feasible.\n\nSemidefinite programming permits stronger relaxations in which matrix variables must be positive semidefinite. For maximum cut, the Goemans–Williamson relaxation assigns a unit vector \\(v_i\\) to each vertex and maximizes\n\n\\[\\sum_{ij}w_{ij}\\frac{1-v_i\\cdot v_j}{2}.\\]\n\nAssignments of vectors to two opposite directions encode ordinary cuts, so this continuous optimum upper-bounds the optimal cut weight.\n\nThe rounding procedure chooses a random hyperplane through the origin and separates vertices according to the sides containing their vectors. If two vectors form an angle \\(\\theta\\), they are separated with probability \\(\\theta/\\pi\\). Comparing this probability with the relaxed edge contribution \\((1-\\cos\\theta)/2\\) gives\n\n\\[\\alpha_{\\mathrm{GW}} =\\min_{0<\\theta\\leq\\pi} \\frac{2\\theta}{\\pi(1-\\cos\\theta)} \\approx0.87856.\\]\n\nConsequently, for nonnegative edge weights the expected output is at least this fraction of the optimum, up to numerical accuracy in solving the relaxation. The improvement over independent random assignment comes from using vector geometry to coordinate the rounding choices."
    },
    {
      "heading": "Metric and geometric approximation",
      "text": "Structural assumptions can change approximability dramatically. A central example is the traveling salesperson problem, which asks for a minimum-cost tour visiting every vertex. In the metric version, costs are symmetric, nonnegative, and satisfy the triangle inequality:\n\n\\[d(u,w)\\leq d(u,v)+d(v,w).\\]\n\nThis inequality permits shortcutting: replacing part of a walk by a direct edge cannot increase its length. Many routing approximation algorithms depend on this operation.\n\nA simple factor-two algorithm begins with a minimum spanning tree. Removing an edge from an optimal tour leaves a spanning tree, so the minimum spanning tree costs no more than the optimum tour. Doubling its edges creates a connected multigraph of even degrees, which has an Eulerian traversal of twice the tree’s cost. Shortcutting repeated vertices produces a Hamiltonian tour without increasing that cost.\n\nChristofides’ algorithm improves the bound to \\(3/2\\). It finds the odd-degree vertices of a minimum spanning tree and adds a minimum-cost perfect matching on those vertices. Their number is even. Shortcutting an optimal tour to visit just these vertices gives a cycle whose alternating edges form two perfect matchings; the cheaper has cost at most half the original optimum. The minimum matching is no more expensive. Combining the tree and matching therefore gives an Eulerian multigraph of cost at most \\(3/2\\) of optimum, and shortcutting yields the required tour.\n\nThe metric assumption is substantive rather than cosmetic. In unrestricted complete graphs, a missing edge of a Hamiltonian-cycle instance can be represented by an extremely expensive edge. An approximation algorithm with any fixed constant ratio could then distinguish graphs with Hamiltonian cycles from graphs without them, by choosing the penalty sufficiently large. Thus unrestricted traveling salesperson admits no polynomial-time constant-factor approximation unless P equals NP, whereas the metric version has strong constant-factor algorithms.\n\nEuclidean geometry provides still more structure. For points in a fixed-dimensional Euclidean space, traveling salesperson admits polynomial-time approximation schemes, established independently through work by Sanjeev Arora and Joseph Mitchell in the 1990s. A characteristic approach partitions space recursively and restricts how a near-optimal tour crosses partition boundaries. Dynamic programming then records a bounded collection of crossing patterns. The analysis shows that an optimal tour can be modified to respect these restrictions with only a small relative increase in length. Fixed dimension and fixed accuracy are essential qualifications: the dependence on those parameters can be substantial even when running time is polynomial in the number of points."
    },
    {
      "heading": "Approximation schemes and numerical scaling",
      "text": "A polynomial-time approximation scheme, or PTAS, is a family of algorithms indexed by an accuracy parameter \\(\\varepsilon>0\\). For minimization it returns a solution of value at most \\((1+\\varepsilon)\\operatorname{OPT}\\); for maximization the usual requirement is at least \\((1-\\varepsilon)\\operatorname{OPT}\\). For every fixed \\(\\varepsilon\\), running time must be polynomial in input size. A running time such as \\(n^{O(1/\\varepsilon)}\\) therefore qualifies, despite becoming expensive as the requested error decreases.\n\nA fully polynomial-time approximation scheme, or FPTAS, satisfies the stronger requirement of polynomial running time in both the input length and \\(1/\\varepsilon\\). This distinction captures the cost of requesting additional accuracy. A PTAS can be efficient for moderate accuracy yet impractical for small \\(\\varepsilon\\); an FPTAS gives a more controlled trade-off. Neither designation by itself guarantees that constants or polynomial degrees are small enough for a particular application.\n\nThe 0–1 knapsack problem provides a standard scaling construction. Each item has a nonnegative profit \\(p_i\\), a weight, and a choice of being included or excluded. The selected items must fit within a capacity. After removing items that cannot fit individually, let \\(P_{\\max}\\) be the largest remaining profit and let \\(n\\) be the number of remaining items. For \\(0<\\varepsilon<1\\), set\n\n\\[K=\\frac{\\varepsilon P_{\\max}}{n}, \\qquad p'_i=\\left\\lfloor\\frac{p_i}{K}\\right\\rfloor.\\]\n\nA dynamic programming algorithm can find an exactly optimal feasible subset for the scaled integer profits, using states that record the minimum weight needed for each total scaled profit.\n\nRounding loses less than \\(K\\) profit per item, hence less than \\(nK=\\varepsilon P_{\\max}\\) across a solution. Since the best feasible individual item implies \\(P_{\\max}\\leq\\operatorname{OPT}\\), the chosen subset has true profit at least\n\n\\[\\operatorname{OPT}-nK \\geq(1-\\varepsilon)\\operatorname{OPT}.\\]\n\nEach scaled profit is at most \\(n/\\varepsilon\\), so the total scaled profit is at most \\(n^2/\\varepsilon\\). The straightforward dynamic program consequently runs in time polynomial in \\(n\\) and \\(1/\\varepsilon\\). If all feasible profits are zero, that trivial case can be handled separately.\n\nThis construction explains how approximation can turn a pseudopolynomial algorithm into a genuinely polynomial one. A pseudopolynomial running time depends on a numerical magnitude, such as total profit, rather than on its binary encoding length. Scaling reduces that magnitude while keeping the induced error under control. By contrast, for strongly NP-hard problems with suitably bounded integer objectives, sufficiently accurate fully polynomial approximation would recover an exact optimum in polynomial time. Such problems therefore generally cannot have an FPTAS unless P equals NP, although they may still admit a PTAS."
    },
    {
      "heading": "Inapproximability and approximation-preserving reductions",
      "text": "Exact NP-hardness alone does not establish the difficulty of approximation. A reduction proving exact hardness might create instances whose optimal values differ by only one unit out of a very large total. A modest relative error could conceal that difference. To establish inapproximability, a reduction must create a controlled gap: one class of inputs has optimum at or below a threshold, while another has optimum above a sufficiently separated threshold.\n\nAn approximation-preserving reduction formalizes how approximate solutions to one problem can be transformed into approximate solutions to another. Different reduction types preserve different properties, such as constant-factor approximability or the existence of approximation schemes. The details matter: a reduction suitable for proving NP-completeness may provide no useful approximation relationship. Objective scaling, auxiliary variables, and forced costs can all change the relevant ratio.\n\nThe PCP theorem, concerning probabilistically checkable proofs, supplies a major source of gap reductions. In one formulation, membership in NP can be verified using logarithmically many random bits while examining only a constant number of proof locations. Translating the verifier’s tests into constraints yields instances where satisfiable cases can be distinguished from cases in which every assignment violates a definite fraction of constraints. This connection makes it possible to rule out algorithms that would approximate certain optimization problems too closely, assuming P differs from NP.\n\nThe class APX contains optimization problems admitting polynomial-time constant-factor approximations, under standard definitions of polynomially verifiable optimization problems. APX-hardness, with an appropriate approximation-preserving reduction, expresses that a problem captures the approximation difficulty of this class. The designation APX-complete additionally requires membership in APX. For standard APX-complete problems, a PTAS would imply P equals NP. Approximation classes therefore distinguish degrees of computational difficulty that ordinary NP-hardness leaves unresolved.\n\nSome thresholds are sharp. For set cover on a universe of \\(n\\) elements, the greedy logarithmic bound is essentially optimal: for every fixed \\(\\varepsilon>0\\), approximation within \\((1-\\varepsilon)\\ln n\\) is NP-hard. Work by Irit Dinur and David Steurer in 2014 established this threshold under the assumption P differs from NP, strengthening earlier results that used stronger complexity assumptions.\n\nOther proposed thresholds depend on additional conjectures. The Unique Games Conjecture, introduced by Subhash Khot in 2002, concerns the difficulty of certain constraint-satisfaction problems. Conditional on this conjecture, the Goemans–Williamson constant is the optimal polynomial-time approximation threshold for maximum cut. Such a statement is stronger in its conclusion but also stronger in its assumption than an ordinary NP-hardness result. Distinguishing proved hardness, conjecture-dependent hardness, and the limitations of a particular technique is essential when describing an algorithm as optimal."
    },
    {
      "heading": "Alternative guarantees and computational models",
      "text": "Not every useful approximation result fits a single multiplicative ratio. In bin packing, items of sizes at most one must be placed into the fewest unit-capacity bins. Algorithms may satisfy a bound such as\n\n\\[A(I)\\leq(1+\\varepsilon)\\operatorname{OPT}(I)+C_\\varepsilon,\\]\n\nwhere \\(C_\\varepsilon\\) is independent of the instance but depends on the requested accuracy. This is an asymptotic approximation guarantee: the additive term becomes relatively insignificant when many bins are needed. It is not the same as a bound of \\((1+\\varepsilon)\\operatorname{OPT}(I)\\) for every instance.\n\nThe difference is forced by small instances. Distinguishing whether a set of items with total size two fits into two bins can encode the partition problem. An absolute approximation ratio strictly below \\(3/2\\) would have to return two bins whenever the optimum is two, because the number of bins is integral. It would therefore distinguish these instances exactly. Additive constants allow strong large-instance guarantees without making such claims about small optima.\n\nBicriteria approximation permits controlled losses in two requirements. A clustering algorithm might obtain a constant-factor objective value while opening more than the specified number \\(k\\) of centers. A scheduling algorithm might meet a target only on machines faster than those in the original instance. These can be mathematically strong results, but their interpretation must include both allowances. A solution that violates a hard constraint is not an ordinary feasible approximation merely because its objective is favorable.\n\nAn online algorithm receives input incrementally and often must make decisions before future requests are known. Its performance is commonly measured by a competitive ratio against an optimal offline solution that knows the complete input. This resembles an approximation ratio algebraically, but the source of difficulty differs: an online algorithm faces missing future information, while an offline approximation algorithm has the entire instance but limited computational resources. An online lower bound can hold even with unlimited computation.\n\nStreaming and distributed models impose further restrictions. A streaming algorithm may have to process a large graph using memory far smaller than the full graph representation; an algorithm in distributed computing may be constrained by communication rounds or message sizes. Approximation quality must then be considered alongside memory, communication, update time, or synchronization. Likewise, parameterized approximation studies trade-offs involving a structural parameter as well as input size. A procedure running in \\(f(k)n^{O(1)}\\) time has a different efficiency guarantee from one running in \\(n^{f(k)}\\). Reporting only the objective ratio can conceal the resource that makes a result significant."
    },
    {
      "heading": "Applications, objective structure, and implementation",
      "text": "Approximation algorithms are particularly useful when the optimization model captures a large discrete planning problem and an exact solution is too expensive to obtain. In scheduling theory, for example, assigning independent jobs to identical machines to minimize the maximum completion time leads directly to makespan approximation. List scheduling has a short, informative proof: if the last-finishing job has processing time \\(p\\), its starting time is at most the total processing time of the other jobs divided by \\(m\\). Thus\n\n\\[C_{\\max}\\leq\\frac{\\sum_jp_j}{m} +\\left(1-\\frac1m\\right)p \\leq\\left(2-\\frac1m\\right)\\operatorname{OPT}.\\]\n\nThe proof depends on both average load and the fact that every feasible schedule must accommodate the longest job.\n\nIn facility location, the objective combines the cost of opening facilities with the cost of serving clients. Constant-factor approximation results often rely on metric connection costs, which support reassignment arguments when facilities are merged, closed, or replaced. In networking, related models choose links that connect required terminals. The Steiner tree problem differs from minimum spanning tree because additional, nonrequired vertices may be used to reduce total connection cost. Such modeling choices affect both the optimum and the applicable algorithms.\n\nObjective structure can be as important as graph or geometric structure. A submodular function expresses diminishing marginal returns: adding an element to a smaller selected set provides at least as much gain as adding it to a larger one. For a normalized, nonnegative, monotone submodular function under a cardinality constraint of at most \\(k\\) elements, greedy selection obtains at least \\(1-1/e\\) of optimum. At each step, the best available marginal gain is at least one \\(k\\)-th of the remaining gap to the optimal value; iterating this inequality gives the bound. Maximum coverage is a prominent example.\n\nThese results connect with feature selection, experimental design, and data summarization, but a guarantee applies only when the stated mathematical assumptions hold. An objective arising in machine learning is not automatically submodular, and approximate or noisy objective evaluation may require a revised analysis. Similarly, a routing guarantee based on symmetric metric distances need not apply to asymmetric travel times or additional time-window constraints.\n\nPractical implementations often combine a provable method with a heuristic. An approximation solution can initialize local improvement, while a relaxation supplies an instance-specific lower bound. If a minimization algorithm finds a solution of cost \\(U\\) and a valid lower bound is \\(L>0\\), then \\(U/L\\) certifies its quality for that instance, potentially much more strongly than the general worst-case theorem. Preserving that certificate requires care with floating-point tolerances, feasibility checks, and termination rules. Rounded numerical output from an optimization solver is not automatically an exact bound; rigorous certificates require validated feasibility or explicit accounting for numerical error."
    }
  ],
  "links": [
    "algorithm",
    "mathematical-optimization",
    "np-hardness",
    "computational-complexity",
    "time-complexity",
    "randomized-algorithm",
    "np-completeness",
    "p-versus-np-problem",
    "greedy-algorithm",
    "integer-programming",
    "linear-programming",
    "semidefinite-programming",
    "knapsack-problem",
    "dynamic-programming",
    "unique-games-conjecture",
    "distributed-computing",
    "machine-learning",
    "heuristic"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}