{
  "id": "normal-vector",
  "title": "Normal Vector",
  "lang": "en",
  "source_lang": "en",
  "summary": "A normal vector is perpendicular to a line, plane, or tangent space, describing local orientation and supporting geometric calculations, surface integration, and physical boundary conditions.",
  "category": "mathematics",
  "subcategory": "geometry",
  "type": "concept",
  "aliases": [
    "Normal",
    "surface normal",
    "perpendicular vector"
  ],
  "generated_at": "2026-11-04",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/normal-vector/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/normal-vector/",
      "json": "https://aiwiki.page/zh/normal-vector.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A normal vector is a nonzero vector perpendicular to a line, a plane, or the tangent space of a geometric object at a specified point. In Euclidean space, perpendicularity is measured by the dot product: a vector \\(\\mathbf n\\) is normal to a tangent direction \\(\\mathbf v\\) when \\(\\mathbf n\\cdot\\mathbf v=0\\). Normal vectors describe the orientation of surfaces and boundaries and are used throughout analytic geometry, differential geometry, physics, and computer graphics. A unit normal is a normal vector of length one."
    },
    {
      "heading": "Definition and normal spaces",
      "text": "Let \\(M\\) be a smooth \\(k\\)-dimensional submanifold of \\(\\mathbb R^n\\), and let \\(p\\in M\\). The normal space at \\(p\\) is the orthogonal complement of its tangent space:\n\n\\[N_pM=(T_pM)^\\perp =\\{\\mathbf n\\in\\mathbb R^n: \\mathbf n\\cdot\\mathbf v=0 \\text{ for every }\\mathbf v\\in T_pM\\}.\\]\n\nIts dimension is \\(n-k\\). Strictly speaking, the normal space contains the zero vector, but a vector used to specify a normal direction is required to be nonzero.\n\nFor a smooth surface in three-dimensional space, the tangent space is a plane and the normal space is a line. Consequently, every normal vector at a given point is a scalar multiple of any other one. There are exactly two unit normals, pointing in opposite directions:\n\n\\[\\widehat{\\mathbf n}=\\frac{\\mathbf n}{\\|\\mathbf n\\|}, \\qquad -\\widehat{\\mathbf n}.\\]\n\nHere \\(\\|\\mathbf n\\|\\) denotes the Euclidean norm. “Normal” does not mean “normalized”: normality specifies perpendicularity, whereas normalization specifies unit length.\n\nA curve in three-dimensional space has a two-dimensional normal space, so perpendicularity to the curve alone does not select a unique normal direction. More generally, a submanifold has a unique normal line only when its dimension is one less than that of the ambient space.\n\nIn an inner-product space, the definition uses the relevant inner product in place of the Euclidean dot product. Normality therefore depends on the metric, not just on the underlying coordinates."
    },
    {
      "heading": "Lines, planes, and hyperplanes",
      "text": "A line in \\(\\mathbb R^2\\) with equation\n\n\\[ax+by=c, \\qquad (a,b)\\ne(0,0),\\]\n\nhas normal vector \\((a,b)\\). A direction vector along the line is \\((-b,a)\\), whose dot product with \\((a,b)\\) is zero.\n\nSimilarly, the plane\n\n\\[ax+by+cz=d\\]\n\nhas normal vector \\((a,b,c)\\), provided its coefficients are not all zero.\n\nIn \\(\\mathbb R^n\\), a hyperplane can be written as\n\n\\[H=\\{\\mathbf x:\\mathbf n\\cdot(\\mathbf x-\\mathbf p)=0\\},\\]\n\nwhere \\(\\mathbf p\\) is a point on the hyperplane and \\(\\mathbf n\\ne0\\) is a normal vector. The normal determines the hyperplane’s orientation, while \\(\\mathbf p\\) determines its position.\n\nDistance and projection\n\nOnce a direction for the unit normal has been chosen, the signed distance from a point \\(\\mathbf q\\) to \\(H\\) is\n\n\\[d_{\\mathrm s}= \\frac{\\mathbf n\\cdot(\\mathbf q-\\mathbf p)}{\\|\\mathbf n\\|}.\\]\n\nIts absolute value is the ordinary perpendicular distance. The orthogonal projection of \\(\\mathbf q\\) onto the hyperplane is\n\n\\[\\mathbf q_H= \\mathbf q- \\frac{\\mathbf n\\cdot(\\mathbf q-\\mathbf p)} {\\|\\mathbf n\\|^2}\\mathbf n.\\]\n\nFor example, the plane \\(2x-y+2z=3\\) has normal vector \\((2,-1,2)\\), of length \\(3\\). The distance from \\((1,0,0)\\) to the plane is therefore\n\n\\[\\frac{|2(1)-0+2(0)-3|}{3}=\\frac13.\\]"
    },
    {
      "heading": "Computing surface normals",
      "text": "Implicit surfaces and gradients\n\nSuppose a surface is described locally by a level equation\n\n\\[F(x,y,z)=c,\\]\n\nwhere \\(F\\) is continuously differentiable. At a point \\(p\\) where the gradient is nonzero,\n\n\\[\\nabla F(p)= \\left( \\frac{\\partial F}{\\partial x}, \\frac{\\partial F}{\\partial y}, \\frac{\\partial F}{\\partial z} \\right)_p\\]\n\nis normal to the surface.\n\nThis follows from the chain rule. If \\(\\mathbf r(t)\\) is a curve lying on the surface and passing through \\(p\\), then\n\n\\[F(\\mathbf r(t))=c \\quad\\Longrightarrow\\quad \\nabla F(\\mathbf r(t))\\cdot\\mathbf r'(t)=0.\\]\n\nThus the gradient is perpendicular to every tangent direction. The corresponding tangent plane at \\(p\\) is\n\n\\[\\nabla F(p)\\cdot(\\mathbf x-p)=0.\\]\n\nFor a sphere centered at the origin,\n\n\\[F(x,y,z)=x^2+y^2+z^2,\\]\n\nthe gradient is \\(2(x,y,z)\\). On a sphere of radius \\(R>0\\), the outward unit normal is therefore\n\n\\[\\widehat{\\mathbf n}(x,y,z)=\\frac{(x,y,z)}{R}.\\]\n\nThe condition \\(\\nabla F(p)\\ne0\\) is essential to this construction. If the gradient vanishes, the formula supplies no direction. This can indicate either a genuinely singular surface or an unsuitable defining equation. For example, \\(F(x,y,z)=z^2\\) defines the smooth plane \\(z=0\\) as a level set, but its gradient vanishes everywhere on that plane.\n\nGraphs of functions\n\nFor the graph of a differentiable function \\(z=f(x,y)\\), use\n\n\\[F(x,y,z)=z-f(x,y).\\]\n\nAn upward-pointing normal vector is\n\n\\[\\mathbf n=(-f_x,-f_y,1),\\]\n\nand the upward unit normal is\n\n\\[\\widehat{\\mathbf n}= \\frac{(-f_x,-f_y,1)} {\\sqrt{1+f_x^2+f_y^2}}.\\]\n\nThe negative of this vector gives the downward unit normal.\n\nParametric surfaces\n\nFor a regular parametrized surface\n\n\\[\\mathbf r(u,v)= \\bigl(x(u,v),y(u,v),z(u,v)\\bigr),\\]\n\nthe partial derivatives \\(\\mathbf r_u\\) and \\(\\mathbf r_v\\) span the tangent plane. Their cross product gives a normal:\n\n\\[\\mathbf n=\\mathbf r_u\\times\\mathbf r_v, \\qquad \\widehat{\\mathbf n}= \\frac{\\mathbf r_u\\times\\mathbf r_v} {\\|\\mathbf r_u\\times\\mathbf r_v\\|}.\\]\n\nRegularity requires the cross product to be nonzero. Reversing the order of the parameters reverses the normal direction.\n\nFor a triangle with noncollinear vertices \\(\\mathbf a,\\mathbf b,\\mathbf c\\), the same construction gives a face normal:\n\n\\[\\mathbf n=(\\mathbf b-\\mathbf a)\\times(\\mathbf c-\\mathbf a).\\]\n\nIts magnitude equals twice the triangle’s area.\n\nSeveral simultaneous constraints\n\nSuppose a submanifold is defined locally by\n\n\\[F_1(\\mathbf x)=\\cdots=F_m(\\mathbf x)=0,\\]\n\nwith linearly independent gradients at the point under consideration. Its normal space is\n\n\\[N_pM= \\operatorname{span} \\{\\nabla F_1(p),\\ldots,\\nabla F_m(p)\\}.\\]\n\nIf \\(J\\) is the Jacobian matrix whose rows are these gradients, then\n\n\\[T_pM=\\ker J, \\qquad N_pM=\\operatorname{im}(J^{\\mathsf T}).\\]\n\nThis connects surface geometry with the null space and image of a matrix."
    },
    {
      "heading": "Normals to curves",
      "text": "For a regular plane curve \\(\\mathbf r(t)=(x(t),y(t))\\), rotating the tangent vector through a right angle gives a normal vector:\n\n\\[\\mathbf n(t)=(-y'(t),x'(t)).\\]\n\nEither sign may be chosen.\n\nFor a space curve, every vector perpendicular to its tangent is normal in the broad sense. A particular choice, the principal normal, is defined using the change of the unit tangent. With arc length \\(s\\) and unit tangent \\(\\mathbf T(s)\\),\n\n\\[\\mathbf N(s)= \\frac{d\\mathbf T/ds}{\\|d\\mathbf T/ds\\|}\\]\n\nwhere \\(d\\mathbf T/ds\\ne0\\). The curvature is\n\n\\[\\kappa(s)=\\left\\|\\frac{d\\mathbf T}{ds}\\right\\|.\\]\n\nThus the principal normal is defined where curvature is nonzero, and points in the direction in which the tangent turns. It should not be confused with an arbitrary member of the curve’s normal space or with the normal to a surface containing the curve."
    },
    {
      "heading": "Orientation and normal fields",
      "text": "Choosing a unit normal at every point of a surface produces a normal vector field. A continuous choice distinguishes two sides of the surface.\n\nFor a smooth embedded surface in \\(\\mathbb R^3\\), the existence of a globally continuous unit normal is equivalent to orientability. An orientable surface admits such a choice; a Möbius strip does not, although normals can still be chosen in sufficiently small neighborhoods.\n\nFor a smooth boundary enclosing a region, the usual convention is the outward unit normal, pointing away from the region’s interior. Changing this choice reverses the sign of oriented flux integrals.\n\nAt edges, corners, and other nonsmooth points, there may be no unique tangent plane and hence no unique classical surface normal. Each adjacent smooth face may have its own normal. In convex optimization, this ambiguity is handled by a normal cone. For a convex set \\(C\\) and \\(p\\in C\\), one convention defines\n\n\\[N_C(p)= \\{\\mathbf v: \\mathbf v\\cdot(\\mathbf x-p)\\le0 \\text{ for every }\\mathbf x\\in C\\}.\\]\n\nAt a smooth boundary point of a full-dimensional convex region, this cone consists of the outward normal ray together with zero."
    },
    {
      "heading": "Applications",
      "text": "Flux and boundary conditions\n\nThe flux of a vector field \\(\\mathbf F\\) through an oriented surface \\(S\\) is a surface integral:\n\n\\[\\Phi=\\int_S\\mathbf F\\cdot\\widehat{\\mathbf n}\\,dS.\\]\n\nThe dot product extracts the component crossing the surface rather than moving tangentially along it. The divergence theorem relates outward flux through a closed boundary to the volume integral of divergence:\n\n\\[\\int_{\\partial V}\\mathbf F\\cdot\\widehat{\\mathbf n}\\,dS = \\int_V\\nabla\\cdot\\mathbf F\\,dV.\\]\n\nNormals also specify boundary data in partial differential equations. The normal derivative of a scalar field is\n\n\\[\\frac{\\partial u}{\\partial n} = \\nabla u\\cdot\\widehat{\\mathbf n}.\\]\n\nPrescribing this derivative is a common form of Neumann boundary condition in a boundary-value problem.\n\nConstrained optimization\n\nAt a regular constrained stationary point of a differentiable function \\(f\\) on a surface \\(g(\\mathbf x)=c\\),\n\n\\[\\nabla f=\\lambda\\nabla g.\\]\n\nThe geometric meaning is that \\(\\nabla f\\) has no component along feasible tangent directions: it lies in the normal space of the constraint. This is the basis of the method of Lagrange multipliers.\n\nComputer graphics and transformations\n\nIn computer graphics, surface normals are used for lighting, visibility tests, reflection, and shading. A polygon has a geometric face normal; interpolated vertex normals may instead be used to give a mesh a visually smooth appearance. Such shading normals need not equal the exact normals of the polygonal surface.\n\nNormals do not generally transform in the same way as positions or tangent vectors. Under an invertible affine transformation\n\n\\[\\mathbf x'=A\\mathbf x+\\mathbf b,\\]\n\na normal transforms, up to nonzero scale, as\n\n\\[\\mathbf n'=A^{-\\mathsf T}\\mathbf n.\\]\n\nIndeed, for a transformed tangent \\(\\mathbf t'=A\\mathbf t\\),\n\n\\[(A^{-\\mathsf T}\\mathbf n)\\cdot(A\\mathbf t) =\\mathbf n\\cdot\\mathbf t=0.\\]\n\nThe resulting vector must usually be normalized. Applying \\(A\\) directly to a normal works for orthogonal transformations and uniform scaling, but generally fails under nonuniform scaling or shear. If orientation is tracked through ordered surface tangents, the sign of the transformation’s determinant must also be accounted for."
    }
  ],
  "links": [
    "tangent-space",
    "euclidean-space",
    "dot-product",
    "analytic-geometry",
    "differential-geometry",
    "orthogonal-complement",
    "norm-mathematics",
    "inner-product",
    "hyperplane",
    "orthogonal-projection",
    "gradient",
    "chain-rule",
    "graph-of-a-function",
    "partial-derivative",
    "cross-product",
    "jacobian-matrix",
    "null-space",
    "vector-field",
    "convex-optimization",
    "divergence-theorem",
    "partial-differential-equation",
    "boundary-value-problem",
    "lagrange-multipliers",
    "computer-graphics"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}