{
  "id": "minkowski-sum",
  "title": "Minkowski Sum",
  "lang": "en",
  "source_lang": "en",
  "summary": "The Minkowski sum combines two sets by adding every vector in one to every vector in the other, providing a fundamental operation in convex geometry, optimization, and computational geometry.",
  "category": "mathematics",
  "subcategory": "geometry",
  "type": "concept",
  "aliases": [
    "Set sum",
    "Vector sum of sets",
    "Minkowski addition"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/minkowski-sum/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/minkowski-sum/",
      "json": "https://aiwiki.page/zh/minkowski-sum.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "The Minkowski sum of two subsets \\(A\\) and \\(B\\) of a vector space is the set of all sums \\(a+b\\), where \\(a\\in A\\) and \\(b\\in B\\):\n\n\\[A+B=\\{a+b:a\\in A,\\ b\\in B\\}.\\]\n\nIt is also commonly written \\(A\\oplus B\\). Named after the mathematician Hermann Minkowski, this operation turns vector addition into an operation on entire sets. It is fundamental in convex geometry and is used to describe geometric offsets, combine uncertainty regions, and compute collision constraints."
    },
    {
      "heading": "Definition and geometric interpretation",
      "text": "For each \\(b\\in B\\), let\n\n\\[A+b=\\{a+b:a\\in A\\}\\]\n\nbe the translation of \\(A\\) by \\(b\\). Then\n\n\\[A+B=\\bigcup_{b\\in B}(A+b).\\]\n\nThus, the Minkowski sum can be visualized by moving a copy of \\(A\\), without rotating it, so that its reference point occupies every position in \\(B\\), and taking the union of all resulting copies.\n\nThe operation is defined whenever the ambient space has an addition law. Its most familiar setting is Euclidean space \\(\\mathbb{R}^n\\), but it also applies to function spaces and more general algebraic structures.\n\nSeveral simple examples illustrate its meaning:\n\n- Intervals: If \\(a\\le b\\) and \\(c\\le d\\), then\n\n\\[[a,b]+[c,d]=[a+c,b+d].\\]\n- Finite sets: If \\(A=\\{0,2\\}\\) and \\(B=\\{0,3\\}\\), then\n\n\\[A+B=\\{0,2,3,5\\}.\\]\n\nDifferent pairs can yield the same sum, so the number of resulting elements need not equal the product of the input sizes.\n- Line segments: Two nonparallel line segments in the plane have a parallelogram as their sum.\n- Balls: Writing\n\n\\[\\overline B_r=\\{x\\in\\mathbb{R}^n:\\|x\\|\\le r\\},\\]\n\none has\n\n\\[\\overline B_r+\\overline B_s=\\overline B_{r+s}\\]\n\nfor \\(r,s\\ge0\\).\n\nA Minkowski sum is neither a union nor a Cartesian product. It combines points through addition rather than merely collecting them or forming ordered pairs."
    },
    {
      "heading": "Algebraic properties",
      "text": "Minkowski addition inherits commutativity and associativity from vector addition:\n\n\\[A+B=B+A, \\qquad (A+B)+C=A+(B+C).\\]\n\nThe singleton containing the zero vector is its identity:\n\n\\[A+\\{0\\}=A.\\]\n\nThe empty set is absorbing:\n\n\\[A+\\varnothing=\\varnothing.\\]\n\nThe operation is monotone under set inclusion:\n\n\\[A\\subseteq A',\\quad B\\subseteq B' \\quad\\Longrightarrow\\quad A+B\\subseteq A'+B'.\\]\n\nIt distributes over unions:\n\n\\[A+(B\\cup C)=(A+B)\\cup(A+C).\\]\n\nFor intersections, only the inclusion\n\n\\[A+(B\\cap C)\\subseteq(A+B)\\cap(A+C)\\]\n\nholds in general.\n\nFor a scalar \\(\\lambda\\), define\n\n\\[\\lambda A=\\{\\lambda a:a\\in A\\}.\\]\n\nThen\n\n\\[\\lambda(A+B)=\\lambda A+\\lambda B.\\]\n\nHowever,\n\n\\[(\\lambda+\\mu)A=\\lambda A+\\mu A\\]\n\nis not generally valid. It does hold when \\(A\\) is a nonempty convex set and \\(\\lambda,\\mu\\ge0\\). In particular, \\(A+A=2A\\) for convex \\(A\\), but not for arbitrary sets. For example, if \\(A=\\{0,1\\}\\), then \\(A+A=\\{0,1,2\\}\\), whereas \\(2A=\\{0,2\\}\\).\n\nFor any linear map \\(T\\),\n\n\\[T(A+B)=T(A)+T(B).\\]\n\nThis allows sums to be projected or transformed by applying the same linear operation to their summands.\n\nMinkowski addition does not generally provide additive inverses. The reflected set\n\n\\[-A=\\{-a:a\\in A\\}\\]\n\nusually satisfies \\(A+(-A)\\ne\\{0\\}\\). Cancellation also fails for arbitrary sets, although it holds for nonempty compact convex sets in Euclidean space."
    },
    {
      "heading": "Convexity and topology",
      "text": "If \\(A\\) and \\(B\\) are convex, their Minkowski sum is convex. Indeed, for \\(a_1,a_2\\in A\\), \\(b_1,b_2\\in B\\), and \\(0\\le t\\le1\\),\n\n\\[t(a_1+b_1)+(1-t)(a_2+b_2) = \\bigl(ta_1+(1-t)a_2\\bigr) + \\bigl(tb_1+(1-t)b_2\\bigr).\\]\n\nThe two parenthesized terms belong to \\(A\\) and \\(B\\), respectively.\n\nMinkowski addition also commutes with taking the convex hull:\n\n\\[\\operatorname{conv}(A+B) = \\operatorname{conv}(A)+\\operatorname{conv}(B).\\]\n\nConsequently, convexifying the inputs and then adding produces the convex hull of the exact sum. It can nevertheless fill holes or gaps present in a nonconvex sum.\n\nTopological properties require some care:\n\n- The sum of two compact sets is compact, because addition is a continuous map on their product.\n- The sum of an open set and any set is open.\n- In Euclidean space, the sum of a compact set and a closed set is closed.\n- Two closed sets can have a sum that is not closed.\n\nFor the last point, consider the closed subsets of \\(\\mathbb{R}^2\\)\n\n\\[A=\\{(x,y):x>0,\\ y\\ge1/x\\}, \\qquad B=\\{0\\}\\times\\mathbb{R}.\\]\n\nTheir sum is\n\n\\[A+B=(0,\\infty)\\times\\mathbb{R},\\]\n\nwhich is not closed. This example also shows that closedness can fail even when both summands are convex."
    },
    {
      "heading": "Support functions and convex bodies",
      "text": "For a nonempty compact convex set \\(K\\subseteq\\mathbb{R}^n\\), its support function is\n\n\\[h_K(u)=\\sup_{x\\in K}\\langle u,x\\rangle,\\]\n\nwhere \\(\\langle u,x\\rangle\\) denotes the Euclidean inner product. Geometrically, it gives the position of a supporting hyperplane with normal \\(u\\).\n\nSupport functions turn Minkowski addition into pointwise addition:\n\n\\[h_{K+L}(u)=h_K(u)+h_L(u).\\]\n\nThey also satisfy\n\n\\[h_{\\lambda K}(u)=\\lambda h_K(u) \\qquad(\\lambda\\ge0).\\]\n\nA nonempty compact convex set is uniquely determined by its support function. Therefore, these identities encode Minkowski addition exactly within that class. They also prove cancellation: if\n\n\\[K+M=L+M,\\]\n\nthen subtracting support functions gives \\(h_K=h_L\\), hence \\(K=L\\).\n\nThis representation connects geometric set addition with convex optimization. Maximizing a linear functional over a sum separates into two independent maximizations:\n\n\\[\\max_{z\\in K+L}\\langle u,z\\rangle = \\max_{x\\in K}\\langle u,x\\rangle + \\max_{y\\in L}\\langle u,y\\rangle.\\]"
    },
    {
      "heading": "Volume inequalities and mixed volumes",
      "text": "Minkowski addition is central to the Brunn–Minkowski inequality. For nonempty compact sets \\(A,B\\subseteq\\mathbb{R}^n\\),\n\n\\[\\operatorname{vol}_n(A+B)^{1/n} \\ge \\operatorname{vol}_n(A)^{1/n} + \\operatorname{vol}_n(B)^{1/n},\\]\n\nwhere \\(\\operatorname{vol}_n\\) is \\(n\\)-dimensional Lebesgue measure.\n\nThe inequality expresses a lower bound on volume growth under set addition. Among convex bodies with nonempty interior, equality occurs precisely when the bodies are translates of positive scalar multiples of one another.\n\nFor compact convex sets \\(K,L\\subseteq\\mathbb{R}^n\\), the volume of their weighted sum is a polynomial:\n\n\\[\\operatorname{vol}_n(K+tL) = \\sum_{j=0}^{n} \\binom{n}{j} V(K[n-j],L[j])\\,t^j, \\qquad t\\ge0.\\]\n\nThe coefficients are mixed volumes; the bracketed notation specifies how many times each set occurs as an argument.\n\nA particularly useful case is the offset \\(K+r\\overline B_1\\). For a planar convex body,\n\n\\[\\operatorname{area}(K+r\\overline B_1) = \\operatorname{area}(K) +r\\,\\operatorname{perimeter}(K) +\\pi r^2.\\]\n\nThis is the planar form of Steiner’s formula, relating Minkowski addition to area and perimeter."
    },
    {
      "heading": "Computation",
      "text": "For finite point sets, a direct method forms all pairwise sums. If the sets contain \\(m\\) and \\(n\\) points, this generates \\(mn\\) candidate points before duplicates are removed.\n\nFor convex polytopes represented by their vertices,\n\n\\[P+Q = \\operatorname{conv}\\{p+q: p\\text{ is a vertex of }P,\\ q\\text{ is a vertex of }Q\\}.\\]\n\nNot every pairwise vertex sum is a vertex of the result.\n\nIn two dimensions, convex polygons admit a more efficient construction. When their vertices are listed in cyclic order, the boundary edge vectors of the sum can be obtained by merging the two cyclic sequences of edge directions. With suitable starting vertices, this takes \\(O(m+n)\\) operations, using big-O notation, and the result has at most \\(m+n\\) edges.\n\nNonconvex inputs are more difficult: their sum can have holes and substantially more boundary structure. A common method decomposes the inputs into convex pieces, computes sums of pairs of pieces, and takes their union. In higher dimensions, output complexity can itself become large, limiting the efficiency of explicit representations."
    },
    {
      "heading": "Applications",
      "text": "Motion planning and collision detection\n\nLet \\(R\\) describe a rigid object relative to a reference point, and let \\(O\\) be an obstacle. With orientation fixed, placing the reference point at \\(t\\) gives the occupied region \\(t+R\\). A collision occurs exactly when\n\n\\[(t+R)\\cap O\\ne\\varnothing,\\]\n\nequivalently,\n\n\\[t\\in O+(-R).\\]\n\nThe sum therefore describes forbidden reference-point positions. This construction is fundamental in robot motion planning. If rotations are also allowed, an additional orientation parameter is required; one fixed Minkowski sum describes only the translational problem.\n\nOffsets and mathematical morphology\n\nThe set \\(A+\\overline B_r\\) is the radius-\\(r\\) offset or thickening of \\(A\\). For nonempty closed \\(A\\subseteq\\mathbb{R}^n\\), it consists exactly of points whose Euclidean distance from \\(A\\) is at most \\(r\\).\n\nIn mathematical morphology, dilation by a structuring set is commonly defined through Minkowski addition. It is used in image processing to expand regions, bridge gaps, and model shape-dependent neighborhoods.\n\nUncertainty and reachable sets\n\nIf uncertain vectors satisfy \\(x\\in X\\) and \\(y\\in Y\\), their possible sums form \\(X+Y\\) when every pair of admissible values is allowed. Dependence constraints can make the actual attainable set smaller.\n\nSimilarly, for a linear dynamical model\n\n\\[x_{k+1}=Ax_k+Bu_k,\\]\n\nthe one-step reachable set is\n\n\\[X_{k+1}=AX_k+BU_k\\]\n\nwhen states and controls range independently over \\(X_k\\) and \\(U_k\\). Here \\(AX_k\\) and \\(BU_k\\) denote linear images of sets."
    },
    {
      "heading": "Related differences and probabilistic sums",
      "text": "The notation \\(A-B\\) often means\n\n\\[A+(-B)=\\{a-b:a\\in A,\\ b\\in B\\}.\\]\n\nThis is different from the Pontryagin difference\n\n\\[A\\ominus B=\\{x:x+B\\subseteq A\\},\\]\n\nwhich describes translations that keep all of \\(B\\) inside \\(A\\). Neither operation is a general inverse of Minkowski addition, and the meaning of “Minkowski difference” varies between fields.\n\nMinkowski addition must also be distinguished from addition of probability distributions. The distribution of a sum of independent random vectors is described by convolution, which accounts for probability weights. A Minkowski sum records possible vector sums without assigning probabilities; for compactly supported independent random vectors, the support of their sum is the Minkowski sum of their supports."
    }
  ],
  "links": [
    "vector-space",
    "euclidean-space",
    "cartesian-product",
    "empty-set",
    "convex-set",
    "linear-map",
    "convex-hull",
    "open-set",
    "closed-set",
    "inner-product",
    "convex-optimization",
    "lebesgue-measure",
    "polygon",
    "big-o-notation",
    "motion-planning",
    "convolution"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}