Motivation
In a low-light detector, a pixel may report a small integer number of detected photons. A constant-variance Gaussian error model can then be misleading: a count with mean one has a different variance from one with mean one hundred. We will build a finite model in which the variance equals the mean, and derive the objective and an algorithm from that model. The calculation applies to a selected idealized imaging problem; it is not a model of a particular detector.
The unknown is the intensity in source pixel , and describes how much of that intensity reaches detector pixel . A known background includes any independently modeled background intensity. For a finite matrix , put . Suppose that, conditional on , the counts are independent Poisson random variables with means . This includes blur, but a calibrated and background are assumptions, not consequences of the observed image.
Definitions
For observed counts , the Poisson probability is . Removing terms independent of from the negative log likelihood gives
The positive background makes every logarithm well-defined, even at . This is a modeling restriction; detectors with zero background can also be modeled, but boundary conventions and zero means require separate care. The gradient and Hessian, calculated on an open neighborhood where all are positive, are
Division is componentwise. A zero observation contributes to the objective and zero curvature; it is still informative and must not be discarded.
Theorem/proposition
Theorem 1.3.1
(Existence and a sufficient condition for uniqueness)
Assume , , and every column of has positive sum. Then has a minimizer on the nonnegative orthant. If the rows of indexed by span , the minimizer is unique. The spanning condition is sufficient, not necessary.
Theorem 1.3.2
(A monotone multiplicative step)
Under the same assumptions, let and start with . Define, for each ,
Every iterate is nonnegative and . This inequality alone does not prove convergence of the iterates to a unique minimizer or tell us when to stop.
The multiplicative form hides a gradient step, which the next description makes explicit.
Concept lensAnalytic
The multiplicative step as a scaled gradient step
Since , the update can be rewritten as
It is a gradient step with coordinate step size . A positive fixed point therefore satisfies in every coordinate, while a coordinate equal to zero never moves. The optimality conditions on the orthant, , and , use the same product that drives the step.
Proof sketch or proof idea
Proof
(Coercivity, then strict convexity)
For each , the scalar function is continuous on , bounded below, and tends to infinity as . Since is nonnegative and every column sum is positive, tends to infinity whenever and . At least one then diverges, and the remaining summands are bounded below. Thus is coercive on the closed orthant; continuity and compactness of its bounded sublevel sets give existence. The Hessian is positive semidefinite. When the positive-count rows span, for every nonzero ,
Along every nonconstant line segment in the orthant the second derivative is therefore positive, so is strictly convex there and has at most one minimizer.
Proof
(Descent by a separable upper bound)
For descent, first suppose the current is strictly positive. Set and ; these nonnegative weights sum to one. For any strictly positive trial vector , concavity of the logarithm yields
where zero-weight terms are omitted and the background term is . Hence, with ,
The right side equals zero at and separates by coordinate. Its minimum over is attained at ; when , interpret the logarithmic term as zero and take the boundary value . By continuity this proves the inequality even if the new iterate has zeros. At later steps, repeat the argument on the positive coordinates, leaving zero coordinates fixed. This establishes monotonicity, not a convergence rate.
Proof X-ray
Where the guarantee comes from
The descent proof is a majorize–minimize argument with one decisive move: Jensen's inequality, applied with the weights that split each predicted count into its source contributions and its background. That inequality turns the coupled term into a sum of one-coordinate terms, giving an upper bound that touches at and separates by coordinate. Minimizing a function that lies above and equals it at can only lower ; it need not minimize in one step.
Each hypothesis has a definite job. Positivity of keeps every positive, so the logarithms and the weights are defined even when coordinates vanish. Positive column sums make each separable problem have a finite minimizer, the update itself. A strictly positive current iterate is needed to form ; zero coordinates are simply left out, which is also why they stay zero forever. None of these assumptions supplies missing information about a blurred image.
Worked examples
Worked example 1.3.3
(One detector and a positive count)
Let , and . Then for . The derivative vanishes at , giving the unique minimizer . Starting at , the update gives ; another step is generally needed. The expected count at the solution is five, not the source intensity two.
Worked example 1.3.4
(A boundary optimum)
With , and , on , so its unique minimizer is zero. The update reaches zero in one step from every positive start. A zero count has not caused a logarithm of zero: it removes the logarithmic term, while the positive background keeps the mean positive.
Worked example 1.3.5
(Two sources, one detector)
Let , and . The objective depends only on , and every nonnegative pair with sum four is a minimizer. The update scales both entries by the same factor , preserving their ratio whenever both are positive. The data cannot identify how the total should be divided. No monotonicity claim repairs that loss of information.
Common mistakes
Counterexample mode
A zero coordinate can hide a better solution
With , and , the unique minimizer is . Starting the multiplicative formula at instead leaves it at zero forever, although . Starting strictly positive avoids this immediate locking; it does not by itself prove a general convergence theorem.
The next two mistakes concern what the data and the iteration can deliver.
Common mistake
(Discarding zero counts)
A zero count still contributes the term , which penalizes predicted counts. With , and , the full objective has the unique minimizer . Dropping the second detector removes the only term containing , so every with becomes a minimizer: the information that the second source is dark is lost.
Common mistake
(Reading monotone descent as identification)
In the two-sources example every pair with sum four is a minimizer, and the update preserves the ratio of positive entries. Starts and therefore lead, if they converge, to the different limits and . Monotonicity is a property of the algorithm; it cannot choose among minimizers the data do not distinguish.
The objective is a likelihood, not a regularizer. A unique maximum-likelihood estimate can still be highly sensitive to noise in a poorly conditioned model. Adding a penalty or stopping early changes the estimator and requires its own analysis. Also, the equality of Poisson mean and variance belongs to this observation model; detector saturation, correlated counts and calibration error can invalidate it.
Quick check
Quick check
(Why is an observation not the same as an absent detector?)
Compare the two contributions to .
Solution · Answer
A zero observation contributes , which penalizes predicted counts at that detector. An absent detector contributes no likelihood term at all.
Exercises
Exercise 1.3.6
(For , , , find the minimizer and the first update from .)
Foundation
Solution(2)
- 1
Solve the stationary equation
gives , the unique minimizer.
- 2
Apply one update
At , and , so .
Exercise 1.3.7
(For , , , find the minimizer and explain why the unconstrained stationary point is not feasible.)
Foundation
Solution(2)
- 1
Locate the stationary point
gives , outside the orthant.
- 2
Use the sign of the derivative
On , , so increases and the minimum is at .
Exercise 1.3.8
(Let be the identity, , and . Find the minimizer and decide whether the theorem's spanning condition holds.)
Core
Solution(3)
- 1
Separate the objective
.
- 2
Minimize each part
The first part is minimized at , so ; the second at . The unique minimizer is .
- 3
Check the spanning condition
Only the first count is positive, and its single row does not span . The minimizer is unique anyway, so the sufficient condition is not necessary.
Exercise 1.3.9
(With , , , start at . Compute the first iterate and describe its limiting total if the iteration converges.)
Core
Solution(3)
- 1
Compute the first iterate
, so both entries are multiplied by : , with total .
- 2
Find the fixed total
Key step
The total obeys , so a positive fixed point satisfies , that is .
- 3
Use the preserved ratio
The ratio stays , giving limiting values under that convergence assumption.