Motivation
A binary pixel labeling represents a region through membership decisions. A continuum alternative represents its boundary as a curve. The two viewpoints share a competition between fitting region intensities and discouraging excessive boundary length, but their mathematics differs. A smooth curve permits a first-variation calculation; a finite graph cut permits a discrete global certificate for a restricted energy. Neither argument automatically proves the other's conclusion.
The main question here is local and precise: if a smooth boundary is displaced slightly outward, how does the energy change? Answering it determines a normal velocity that decreases energy as long as a sufficiently smooth evolution exists. It does not establish that such an evolution remains smooth forever, reaches a global optimum, or resolves changes of topology. A shrinking circle will demonstrate why that qualification is substantive.
We assume multivariable differentiation, change of variables, elementary curve geometry, and scalar differential equations. The transport and length formulas used in the proof are developed below. General bounded-variation compactness, weak curvature evolution, and viscosity solutions are outside this note.
Definitions
Definition 10.1.1
(Domain, outward normal, and curvature)
Let be bounded with smooth boundary. Let have closure contained in and a simple closed boundary . Write for the outward unit normal and define signed curvature using its normal extension near the curve. A circle of radius has . For a counterclockwise arclength parametrization with tangent , .
Definition 10.1.2
(Region-fitting energy)
Assume , , and . With fixed constants define
Both regions are assumed to have positive area when their constants are optimized. Their unique least-squares constants are their means. We can therefore study either fixed constants or the reduced energy obtained after substituting both means. This choice changes the model and must be reported.
Definition 10.1.3
(Normal variation and normal gradient)
A normal variation has boundary for a sufficiently smooth scalar and small . Smallness ensures an embedded curve inside the same domain. The normal gradient of an energy is the function for which its first variation is . This uses the inner product on normal curve velocities, not an metric on arbitrary defining functions.
Definition 10.1.4
(Level-set representation)
A level-set representation uses a function negative inside , zero on , and positive outside, with near the boundary. Then on the interface. The defining function is not unique; the curve is its zero set.
Theorem/proposition
Theorem 10.1.5
(First variation with fixed or optimized region means)
Under these hypotheses, for fixed constants the normal energy gradient is
The same formula holds for the reduced energy when the constants are the means of the current regions, each of positive area. In particular, stationarity under all smooth normal variations requires ; it does not imply global minimality.
Theorem 10.1.6
(Conditional energy descent and the level-set sign)
Suppose a sufficiently smooth family of embedded curves exists on a time interval, stays inside , keeps both region areas positive, and moves with outward normal speed . Use either fixed constants or their current means consistently. Then
For a spatial, temporal defining function with nonzero gradient near the interface, its boundary motion satisfies
This is a consequence conditional on smooth existence, not a theorem supplying a global level-set solution.
The positive sign in the defining-function equation and the negative sign in outward velocity are compatible. A growing positive defining function at a fixed location can move its zero set inward. Checking that sign on a shrinking circle is safer than memorizing a formula without a normal convention.
Proof sketch or proof idea
Proof
(Length, transport, and the vanishing mean derivative)
Differentiate the length using the original arclength parameter. Since , the derivative of its speed at zero is . Integration gives . The outward normal convention makes an expanding circle have positive length derivative, as it should.
For a fixed continuous density , use tubular coordinates . The area element is near the boundary. The signed strip added by the variation has running from zero to . Hence
The formula also handles negative through a signed integral. Uniform continuity in the compact tube makes the remainder negligible after division by . Setting gives the area variation. The complementary region loses exactly the same strip, so its integral has the opposite sign. Applying these formulas to the two fidelity densities proves the displayed .
For a positive-area region, completing the scalar square shows its mean uniquely minimizes the squared fitting integral. More explicitly, if , then the partial derivative of the fitting term with respect to is . The transport formulas show that the means vary differentiably while areas remain positive. In the chain rule for the reduced energy, every mean derivative is multiplied by this zero partial derivative. Therefore only the same boundary variation remains. This explains why the means may be updated without adding an omitted force term, and why the positive-area condition is needed.
Proof X-ray
Why no derivative of the mean appears in the force
The first variation uses each hypothesis at a definite place. The boundary supplies curvature and a tubular neighborhood, so the area element is available. Continuity of the density, here with , gives the uniform estimate that removes the remainder after division by . Neither step says anything about the constants.
The key move concerns the reduced energy. The mean does change when the region moves, but it enters only through a scalar at which the fitting term is stationary, so in the chain rule its derivative is multiplied by zero. Positive area is used exactly here: to define the mean and to differentiate the quotient . An arbitrary prescribed update rule for the constants has no such stationarity, and a mean-derivative term would then survive. If jumped across , the density would not be continuous in the tube, and the transport formula would need one-sided traces.
The second theorem then needs only the first variation and the chain rule along the interface.
Proof
(Dissipation and the level-set sign)
For a smooth evolution substitute into the first variation to obtain the dissipation identity. It is nonincreasing, not necessarily strictly decreasing: a stationary curve has zero derivative. For the level-set equation, a point on the moving boundary obeys . Tangential velocity contributes zero and the normal contribution is . This proves the sign. A smooth extension of away from the interface can define a local evolution equation for , but only its interface values determine this geometric law. Different extensions do not, by themselves, establish global weak uniqueness. Nor do they specify which curve or region is a legitimate continuation after a singularity; those questions belong to a separate PDE theory.
Worked examples
Worked example 10.1.7
(A shrinking circle reaches a genuine singularity)
If the image and both fitted constants are the same constant, take and . Only length contributes. A circle satisfies , hence
Its energy is , with derivative . The dissipation integral gives exactly the same value because and length is . At the extinction time the radius vanishes and curvature diverges. The classical embedded-curve hypotheses fail, so our proof cannot simply continue through that time or declare a unique topology-changing solution.
Worked example 10.1.8
(A smooth image with an exact attracting radial boundary)
Let be a disk large enough to contain the evolving circle, let , fix , and take both fidelity weights one with . On radius , the force is , so . For the exact solution is
It approaches . The energy differs from a radius-independent constant by , whose positive-radius minimum is at that radius. This is a solved radial family with fixed constants; it is not a global uniqueness theorem over every region.
Worked example 10.1.9
(Optimized means change the radial force)
Keep but let have radius two and optimize both means for a circular region of radius . Polar integration gives and . With unit weights and no length or area penalty, the force is . The reduced energy is
At , the means are , outward speed is two, and energy derivative is , matching . Within this radial family the minimum occurs at , where the means are one and three. Updating means therefore changes the equilibrium from the preceding fixed-constant example.
Concept lensGeometric
The curve and its defining function use different metrics
Replacing by , where is strictly increasing and smooth with and , preserves the boundary and its normal. The geometric normal speed remains the same when the defining function is transformed consistently. By contrast, taking a gradient in the space of all defining functions depends on the chosen metric and regularized indicator. It need not produce this exact curve- velocity. An implicit representation is useful, but representation alone is not a PDE well-posedness theorem.
In a discrete alternating scheme, recomputing exact means on regions of positive area cannot increase the fitting energy for fixed regions. If a region has zero area, retaining its previous constant is a defined policy because its own integral is zero. This policy does not make its mean mathematically defined. In the smooth continuum argument, extinction or a zero-area region ends the stated hypotheses; stop that classical evolution or supply a separately justified continuation theory.
Common mistakes
Counterexample mode
A stationary circle can be unstable
Take a constant zero image, fixed levels zero inside and one outside, unit fidelity weights, , and . Then and . Radius is stationary, but a slightly larger circle grows and a smaller one shrinks. Its radius-dependent energy is , which has a maximum there. Solving the Euler condition is therefore not a proof of local or global minimality. If both means were reoptimized for this constant image, they would instead both be zero and the force would change.
An image discontinuity exactly on the interface can invalidate the continuous-density transport argument; one-sided traces need a separate formulation. Explicit time steps need stability and geometry checks even when the continuous law dissipates energy. Reinitializing a defining function should preserve the intended zero set, but an inaccurate numerical operation can move it. None of these issues is resolved by writing a curvature formula alone. If the curve touches the fixed domain boundary, a separate contact condition is needed; the first variation of an interior curve does not determine that boundary behavior.
Common mistake
(Reversing the level-set sign)
Writing to match the minus sign in reverses the motion. For the length-only circle with , it gives , so : the circle grows and its energy increases. The correct law comes from with , which gives .
Summary
Smooth normal variations derive perimeter, area, and regional fidelity forces with a consistent outward sign. Optimal region means contribute no extra first-order term while both regions have positive area. The negative shape gradient dissipates energy conditionally on smooth existence, and a regular level-set representation encodes that velocity. Exact circle examples exhibit attraction, changed equilibria, and finite-time singularities, making the boundary of the proved theory explicit.
Exercises
Exercise 10.1.10
(Take a constant image and set both fitted constants equal to that image value, with and . Solve the circle evolution from radius , state its smooth time interval, and check the dissipation identity.)
Foundation
Solution(3)
- 1
Compute the force and solve
Only the area term remains, so and . Hence for .
- 2
Compare both sides of the dissipation identity
Energy is plus a constant, with derivative . The negative integral of along the circle is , the same value.
- 3
Locate the end of the hypotheses
At the radius vanishes, and the embedded positive-radius hypothesis ends.
Exercise 10.1.11
(In the fixed-level radial example , now set . At radius one, compute , outward speed, and instantaneous energy derivative.)
Foundation
Solution(2)
- 1
Add the length and data forces
Curvature is one and the data force is . Thus .
- 2
Read off the speed and energy rate
, and the energy derivative is . At this radius length and data both drive the boundary inward.
Exercise 10.1.12
(In the optimized-mean radius-two domain example, compute both means and the force at . What changes if a positive length penalty is added?)
Core
Solution(3)
- 1
Evaluate the means and the data force
and ; the data force is zero.
- 2
Interpret the zero force
Key step
With no other terms, vanishes identically on this circle, so it is stationary under all smooth normal variations; the radial minimality claim is separate.
- 3
Add a length penalty
A positive length weight adds to , so the outward speed is there. The circle moves inward and is no longer stationary for the modified energy.
Exercise 10.1.13
(Use to verify the level-set sign on a circle. Explain why its vanishing gradient at the center does not invalidate the regular-interface argument before extinction but does matter at extinction.)
Core
Solution(2)
- 1
Substitute into the level-set law
On the interface and . The equation therefore gives , the outward speed .
- 2
Locate the critical point
Before extinction the center is away from the interface, so a tubular neighborhood of the curve has nonzero gradient. At radius zero the interface reaches that critical point and is no longer a regular closed curve, so the smooth proof ceases to apply.