Why Symmetry Helps in an Extrema Search
An absolute-extrema search can produce many candidate points, especially when a region has several boundary pieces. If a reflection or rotation maps the region onto itself and leaves the objective unchanged, some of those points occur in pairs with equal function values. One calculation can then account for both members of a pair.
The benefit is a shorter candidate search, not a different optimization problem. A valid reduction must preserve every possible absolute-extremum value. It must also retain distinct candidate values: symmetry can show that two points have the same value, but it does not show that all candidates share one value.
Symmetry of a Region and Its Objective
A reflection or rotation of the plane can be written in affine form as \(T(\mathbf{x})=Q\mathbf{x}+\mathbf{b}\), where \(Q\) is an orthogonal matrix, so \(Q^{\top}Q=I\), and \(\mathbf{b}\) is a fixed vector. The vector \(\mathbf{b}\) allows the transformation to be centered away from the origin. For example, reflection across the line \(x=1\) sends \((x,y)\) to \((2-x,y)\).
A point and all its images under the available symmetries form its symmetry orbit. For a reflection, an orbit consists of one point on the reflection axis or a pair of reflected points. For a half-turn about a center, an orbit consists of one point at the center or a pair of opposite points. Every point in an orbit has the same objective value when \(f(T(\mathbf{x}))=f(\mathbf{x})\).
When Reducing to Representatives Is Valid
To make a complete search, choose a subset \(S\subseteq R\) that contains at least one point from every symmetry orbit in \(R\). Such a set is a set of orbit representatives. If an absolute maximum or minimum is attained on \(R\), its value is also attained at a representative: the representative has the same objective value as the original point.
The attainment condition matters. Without it, there may be no point whose value is the absolute maximum or minimum to transfer to a representative. Compactness and continuity are a common sufficient condition: a continuous function on a compact region attains both extrema.
Why Symmetry Preserves Interior Critical Points
Suppose \(f\) is differentiable on an open set containing an interior point \(\mathbf{x}\), and the symmetry identity holds in a neighborhood of that point. Write \(T(\mathbf{x})=Q\mathbf{x}+\mathbf{b}\), where \(Q^{\top}Q=I\). The chain rule applied to \(f(T(\mathbf{x}))=f(\mathbf{x})\) gives
Since \(Q\) is invertible, the equation implies that \(\nabla f(\mathbf{x})=\mathbf{0}\) if and only if \(\nabla f(T(\mathbf{x}))=\mathbf{0}\). Thus a symmetry maps interior critical points to interior critical points. The affine term \(\mathbf{b}\) does not affect this conclusion: differentiating \(Q\mathbf{x}+\mathbf{b}\) with respect to \(\mathbf{x}\) gives \(Q\).
Verify \(T(R)=R\) and \(f(T(\mathbf{x}))=f(\mathbf{x})\) on the region.
Use a half-region, sector, or another subset that meets every orbit, and include its symmetry axes or fixed points.
Check interior critical points and the boundary pieces of the reduced region, including any cut introduced by the reduction.
Pair points with equal values, but compare every distinct value before identifying the absolute extrema.
Worked Example: Reflection Symmetry on a Rectangle
Worked Example: Reducing a Rectangular Search by Reflection
Let \(R=[-1,1]\times[-1,1]\) and \(f(x,y)=x^2+2y^2+2y\). The reflection \(T(x,y)=(-x,y)\) preserves the rectangle. Also,
So it is enough to search \(0\le x\le1\), while including the axis \(x=0\). Since \(R\) is compact and \(f\) is continuous, its absolute extrema are attained.
First find interior critical points. The partial derivatives are \(f_x=2x\) and \(f_y=4y+2\). Solving \(2x=0\) and \(4y+2=0\) gives \(x=0\) and \(y=-\tfrac12\). Its value is
On the horizontal edges \(y=-1\) and \(y=1\), the restrictions are respectively \(f(x,-1)=x^2\) and \(f(x,1)=x^2+4\). Their derivatives with respect to \(x\) are \(2x\), so the edge critical point is \(x=0\). The horizontal-edge candidate values are \(0,1\) on \(y=-1\) and \(4,5\) on \(y=1\), including the endpoints \(x=\pm1\).
On either vertical edge \(x=\pm1\), the restriction is \(1+2y^2+2y\). Its derivative is \(4y+2\), which vanishes at \(y=-\tfrac12\). At that point the value is
The vertical-edge endpoints have values \(1\) at \(y=-1\) and \(5\) at \(y=1\). Reflection pairs the points on the left and right edges; both members give the same values. The distinct candidate values are therefore \(-\tfrac12,0,\tfrac12,1,4,5\). Comparing them shows that the absolute minimum is \(-\tfrac12\), at \((0,-\tfrac12)\), and the absolute maximum is \(5\), at \((1,1)\) and \((-1,1)\).
As a check, write \(f(x,y)=x^2+2(y+\tfrac12)^2-\tfrac12\). On the rectangle, \(x^2\ge0\) and the squared term is nonnegative, so \(f\ge-\tfrac12\), with equality at \((0,-\tfrac12)\). Also \(x^2\le1\) and \(y+\tfrac12\le\tfrac32\), while \(y+\tfrac12\ge-\tfrac12\), so \(2(y+\tfrac12)^2\le\tfrac92\) and \(f\le1+\tfrac92-\tfrac12=5\). Equality occurs at \(y=1\) and \(x=\pm1\), confirming both results.
Worked Example: Half-Turn Symmetry on a Disk
Worked Example: Pairing Boundary Candidates by a Half-Turn
Consider the closed unit disk \(R=\{(x,y):x^2+y^2\le1\}\) and \(f(x,y)=x^2+y^2+xy\). The half-turn \(T(x,y)=(-x,-y)\) preserves the disk, and
In the interior, \(f_x=2x+y\) and \(f_y=x+2y\). Setting both equal to zero gives \(y=-2x\), followed by \(x+2(-2x)=-3x=0\). Thus \(x=y=0\), with value \(f(0,0)=0\).
Parametrize the boundary by \(x=\cos t\), \(y=\sin t\), for \(0\le t\le2\pi\). The restricted function is
Its derivative is \(\cos(2t)\). Setting it equal to zero gives \(2t=\frac{\pi}{2}+k\pi\), or \(t=\frac{\pi}{4}+\frac{k\pi}{2}\), for \(k=0,1,2,3\) in the interval. At \(t=\frac{\pi}{4}\), the point is \((\frac{\sqrt2}{2},\frac{\sqrt2}{2})\) and the value is \(1+\frac12=\frac32\). Its half-turn partner has the same value. At \(t=\frac{3\pi}{4}\), the point is \((-\frac{\sqrt2}{2},\frac{\sqrt2}{2})\) and the value is \(1-\frac12=\frac12\); its partner also has value \(\frac12\).
The complete distinct candidate values are \(0,\frac12,\frac32\). Hence the absolute minimum is \(0\) at the origin and the absolute maximum is \(\frac32\) at \((\frac{\sqrt2}{2},\frac{\sqrt2}{2})\) and its opposite point. Substitution verifies the maximum value: \(x^2+y^2=1\) and \(xy=\frac12\), so \(f=1+\frac12=\frac32\). The boundary minimum candidate is \(\frac12\), which is greater than the interior value \(0\).
Worked Example: Quarter-Turn Symmetry on a Square
Worked Example: Reducing a Square to One Quarter
Let \(R=[-1,1]\times[-1,1]\) and \(f(x,y)=x^2+y^2+x^2y^2\). A quarter-turn \(T(x,y)=(-y,x)\) maps the square onto itself. Moreover,
Every orbit meets the first-quadrant square \(0\le x\le1,\ 0\le y\le1\). In the interior, the partial derivatives are \(f_x=2x(1+y^2)\) and \(f_y=2y(1+x^2)\). Since \(1+x^2\) and \(1+y^2\) are positive, both derivatives vanish only when \(x=y=0\); the origin is on the boundary of the reduced square, not in its interior.
On the edge \(x=1\), the restriction is \(f(1,y)=1+2y^2\). Its derivative is \(4y\), so the edge critical point \(y=0\) has value \(1\); the endpoints \(y=0,1\) have values \(1,3\), respectively. On the edge \(y=1\), the restriction is \(f(x,1)=1+2x^2\), giving the corresponding values \(1\) at \(x=0\) and \(3\) at \(x=1\).
The two cut edges are parts of the reduced region's boundary and must also be checked. On \(x=0\), \(f(0,y)=y^2\), which has endpoint values \(0,1\). On \(y=0\), \(f(x,0)=x^2\), which has endpoint values \(0,1\). The distinct candidate values on the reduced square are \(0,1,3\). Thus the minimum is \(0\) at the origin and the maximum is \(3\) at \((1,1)\). Quarter-turn symmetry carries \((1,1)\) to the other three corners, all with value \(1+1+1=3\). Substitution at the origin gives \(f(0,0)=0\), confirming the minimum.
A Common Mistake: Symmetry of the Region Is Not Enough
A region may be symmetric even when its objective is not. For example, the unit disk is preserved by the half-turn \(T(x,y)=(-x,-y)\), but the objective \(f(x,y)=x\) does not satisfy \(f(T(x,y))=f(x,y)\). In fact,
A tempting but incorrect calculation would pair \((1,0)\) with \((-1,0)\), discard one point, and claim the same candidate value represents both. The displayed values show that this is false. The correct check is to verify objective invariance before pairing candidates. Here, the disk is symmetric but \(f(-x,-y)=-f(x,y)\), not \(f(x,y)\); the symmetry therefore gives no equal-value reduction.
A second error is to calculate one representative value and then omit other distinct candidates. In the rectangle example, reflection paired points across the \(y\)-axis, but it did not make the values \(0,\frac12,1,4,\) and \(5\) equal. Pairing reduces repeated work; it does not replace the final comparison.
What to Keep in a Reduced Search
When a symmetry reduction cuts a region in half or restricts a circular search to a sector, the cut itself may become part of the reduced region's boundary. Include it in the reduced search. Points on a symmetry axis are fixed by a reflection, and points at a rotation center are fixed by rotations; these points may form one-point orbits and must not be lost.
Likewise, boundary pieces must be paired only when the symmetry actually maps one piece to the other. A symmetry can exchange edges, arcs, endpoints, or junctions, but any unpaired piece remains part of the candidate search. Once the candidates are found, record their function values, combine equal values from paired points, and compare every distinct value. This preserves the original absolute-extrema problem while avoiding duplicated calculations.
Check Your Understanding
For each question, distinguish what the symmetry proves from what still needs to be checked.
- What two invariance conditions must be verified before using a reflection to pair candidates?
- Why does an orbit-representative reduction preserve absolute-extremum values when extrema are attained?
- For \(T(\mathbf{x})=Q\mathbf{x}+\mathbf{b}\), what role does the affine term \(\mathbf{b}\) play in the gradient calculation?
- Why must a symmetry axis or a cut boundary be checked when searching a reduced region?
- The disk is preserved by \((x,y)\mapsto(-x,-y)\). Does that alone justify pairing points for \(f(x,y)=x\)? Explain.